Nicolas Knotzer

# **Product Recommendations in E-Commerce Retailing Applications**

Nicolas Knotzer

### **Product Recommendations in E-Commerce Retailing Applications**

The book deals with product recommendations generated by information systems referred to as recommender systems. Recommender systems assist consumers in making product choices by providing recommendations of the range of products and services offered in an online purchase environment. The quantitative research study investigates the influence of psychographic and sociodemographic determinants on the interest of consumers in personalized online book recommendations. The author presents new findings regarding the interest in recommendations, importance of product reviews for the decision process, motives for submitting ratings as well as comments, and the delivery of recommendations. The results show that opinion seeking, opinion leading, domain specific innovativeness, online shopping experience, and age are important factors in respect of the interest in personalized recommendations.

Nicolas Knotzer studied business administration with the focus on information systems, management control and project management. From 2001 to 2006 he joined the Institute for Management Information Systems at the Vienna University of Economics and Business Administration. The author received his doctoral degree in 2006.

Nicolas Knotzer - 978-3-631-75452-8 Downloaded from PubFactory at 01/11/2019 04:40:47AM

via free access

Product Recommendations in E-commerce Retailing Applications

### **Forschungsergebnisse der Wirtschaftsuniversitat Wien**

Band 17

Nicolas Knotzer

### **Product Recommendations in E-commerce Retailing Applications**

#### **Blbllographlc Information published by the Deutsche Natlonalblbllothek**

The Deutsche Nationalbibliothek lists this publication in the Deutsche Nationalbibliografie; detailed bibliographic data is available in the internet at <http://www.d-nb.de>.

Open Access: The online version of this publication is published on www.peterlang.com and www.econstor.eu under the international Creative Commons License CC-BY 4.0. Learn more on how you can use and share this work: http://creativecommons. org/licenses/by/4.0.

This book is available Open Access thanks to the kind support of ZBW – Leibniz-Informationszentrum Wirtschaft.

> Cover design: Atelier Platen according to a design of the advertising agency Publique.

University logo of the Vienna University of Economics and Business Administration. Printed with kindly permission of the University.

Sponsored by the Vienna University of Economics and Business Administration.

> ISBN 978-3-631-75452-8 (eBook) ISSN 1613-3056 ISBN 978-3-631-56622-0

© Peter Lang GmbH lnternationaler Verlag der Wissenschaften Frankfurt am Main 2008 All rights reserved.

All parts of this publication are protected by copyright. Any utilisation outside the strict limits of the copyright law, without the permission of the publisher, is forbidden and liable to prosecution. This applies in particular to reproductions, translations, microfilming, and storage and processing in electronic retrieval systems.

Printed in Germany 1 2 3 4 5 7

www.peterlang.de

### **Abstract**

The book deals with product recommendations generated by information systems referred to as recommender systems. Recommender systems assist consumers in making product choices by providing recommendations of the range of products and services offered in an online purchase environment. The quantitative research study investigates the influence of psychographic and sociodemographic determinants on the interest of consumers in personalized online book recommendations. The book starts with an introductory chapter that sets out the research goal and presents the organization of the work. In Chapter 2 the author establishes working definitions, introduces a general classification and presents application models and business goals of recommender systems. Further, a model of the consumer decision process and the relevancy of virtual commwuties for recommendation purposes is described. Chapter 3 reviews functional aspects of recommender systems. Input and output data, measurement scales for preference elicitation as well as recommendation methods a.re elaborated in detail. Chapter 4 describes the research model, the hypothesis, and the methodology. The results of the empirical study a.re presented in Chapter 5. Structural equation modeling and regression analysis a.re used to verify the hypotheses. The author presents new findings regarding the interest in recommendations, importance of product reviews for the decision process, motives for submitting ratings and comments, and the delivery of recommendations. In particular the results show that ophuon seeking, opinion leading, domain specific innovativeness, online shopping experience, and age are important factors in respect of the interest in online recommendations. The book clos«:>,s with an chapter that summarizes the results, shows limitations of the research conducted, and points out directions for further research.

## **Acknowledgements**

Foremost, I would like to thank my supervisors Prof. Hans Robert Hansen and Prof. Gustaf Neumann. I am very grateful for the discussions, suggestions, and insights that helped me to complete this doctoral dissertation. In the last four years Prof. Hansen has given me a very productive and agreeable working environment and helped me to grow as a researcher as well as a person.

Further, I would like to thank my colleagues at the Institute for Management Information Systems and the Institute for Information Systems and New Media, especially Maria Madlberger, Bernd Simon, Horst 1\·eiblmaier, and Christina Stahl.

Last, but not least, my thanks go to my family and Katrin. I am very indebted to you. Without your support and patience the writing of this book would have not been possible.

### **Contents**


Downloaded from PubFactory at 01/11/2019 04:40:47AM via free access




#### **153**


## **List of Figures**





## **List of Tables**





### **Chapter 1**

## **Introduction**

Information systems that *assist consumers in the buying decision process* are recognized to be one of the most promising appliances in e-commerce environments [SP02, SV99]. In this context *recommender systems* support the consumer in this process by providing *recommendations* of products and services to help customers find products to purchase [SKR0l]. Recommender systems aid the consumer by reducing information overload, providing personalized product information, ranking products according to the individual user's preferences, providing community critiques, and summarizing community opinion [Run00]. Hence, recommender systems represent interesting opportunities for e-commerce vendors to deliver value-added services to the customer. Recommender systems ideally assist e-commerce vendors in turning new and infrequent visitors of the web-site into buyers, building credibility through community inputs, inviting customers back, improving cross sales, and building long term relationships [SKR0l].

#### **1.1 Research Goal**

The majority of research literature regarding recommender systems deals with this topic from the viewpoint of computer science. The focus is on the w1 derlying algorithms for genera.ting recommendations [KSS03, SKKR00, BS97,

Bur02, SVA97, Run00]. The existing research in respect of the marketing perspective (e.g. the influence of recommendations on consumers decisions) is still scarce (SN04, HK04, HM03, CLA+o3, HT00]. Therefore, the author has decided to address the research field "recommender systems" from a marketing perspective.

As examined by Katz and Lazarsfeld in their classic article "Personal Influence: The Part Played by People in the Flow of Mass Communications" interpersonal communication (i.e. "word-of-mouth") is a very important factor in the buying decision proceas [KL55]. By providing community critiques and summarizing community opinion, recommender systems may be used to facilitate interpersonal communication between customers. In this context, the question arises which psychographic and sociographic factors determine the interest in recommendations as well as the interest in word-of-mouth.

Thus, the thesis strives to *identify the underlying psychographic and sociodemographic determinants* that define: (1) the consumer's interest in *personalized recommendations,* (2) the consumer's interest in participating actively in virtual communities of transaction located at online purchase environments by *submitting product-related ratings and comments,* and (3) the consun1er's interest in *product-related opinions of other consumers* in virtual communities.

The author tries to address this question in the context of online book recommendations. This product class was chosen for the following reasons:


Further, the following research topics are addressed by means of exploratory research:


In the thesis, a quantitative approach for the examination of the research questions is applied. Consumers were asked to answer a standardized web- based questionnaire regarding recommendations and product-related word-ofmouth. The research model is verified by the application of *structural equation modeling* and *regression analysis.* 

### **1.2 Contents and Organization**

This doctoral dissertation is divided into *six chapters* that review relevant marketing and computer science literature, introduce the underlying theory and hypotheses, describe the study methodology, and present the results as well as conclusions, implications and limitations.

In the following chapter, recommender systems a.re examined from a *marketing perspective.* At first, the working definitions of the thesis and a taxonomy of recommender systems are established. The next section takes a look at business goals of recommender systems and introduces the corresponding application models. The following section deals with the consumer decision process and how this process may be influenced by recommender systems.

In the last section of Chapter 2 virtual communities and their relevancy for recommendation purposes are addressed.

Chapter 3 reviews *functional aspects* of recommender systems. In the first section of the chapter, input and output data of recommender systems are illustrated. Further, approaches to provide supplementary explanations (i.e. why certain products are recommended) are investigated. In the next section, different *statistical measurement scales* for the elicitation of preferences are introduced. Additionally, the information delivery aspects of recommender systems are highlighted. The chapter concludes with a section that gives a detailed review of recommendation methods and their corresponding advantages and disadvantages.

Chapter 4 deals with the research model, the hypothesis and the methodology of the thesis. In this connection, the problem statement of the thesis is elaborated and the research questions are introduced. Furthermore, the underlying theoretical framework and the hypothesis derived from the research questions and the framework are described. The final section of this chapter focuses on the methodological aspects of the thesis and introduces the research design.

In Chapter 5 the results of the study are set out. In the first step, the descriptive results are presented (i.e. results that are not related to the hypotheses and the research model respectively). In the next section the verification of the research model is conducted. This section is structured in three parts: ( 1) hypotheses regarding psychographic factors that a.re verified using stmctural equation modeling, (2) hypotheses regarding psychogra.phic factors that a.re scrutinized using a regression model, and {3) demographic hypotheses.

Chapter 6 reviews the dissertation's implications and limitations. In addition, directions for further research a.re outlined.

## **Chapter 2**

## **Recommender Systems** - **Definition, Classification, and Marketing Perspectives**

This chapter deals with recommender systems from a marketing perspective. First, *working definitions* of the book are established. Section 2.2 introduces a *general classification* of recommender systems. In this taxonomy recommender systems are classified along user adaption (i.e. personalization aspects), mode of information delivery, method of data acquirement, and recommendation methods. Furthermore, requirements of an "ideal" recommender systems are presented.

Section 2.3 elaborates on different *application models* of recommender systems. These application models are tied to specific *business goals.* These goals are: (1) tw-ning visitors into buyers, (2) building credibility through community, (3) inviting customers back, (4) cross-selling, and (5) building long term relationships. Application models and their corresponding business goals are exemplified by illustrating use cases in different companies or research institutions on the Internet.

Section 2.4 deals with the *consumer decision process.* As recommender systems are designed to assist the consumer in this process, understanding the consumer decision process is an important issue. A phase model of the consumer decision

process as proposed by Blackwell et al. is set out [BMEOl]. In this model all fundamental constructs of consumer behavior in regard to the decision process are integrated and interrelated. It includes the following seven phases: (1) need recognition, (2) search for information, (3) pre-purchase evaluation of alternatives, (4) purchase, (5) consumption, (6) post-consumption evaluation, and (7) divestment. The section describes, how consumers may be supported in these phases by recommender systems.

The last section of this chapter addresses virtual communities and their relevancy for recommendation purposes. Characteristics and benefits of virtual communities of transaction (i.e. virtual communities, whose focus is on the transaction of products and services) are presented. Further, the importance of network effects in virtual communities is highlighted. The section ends with a description of community building aspects.

### **2.1 Working Definitions**

*Recommender systems* are information systems, that assist the user in making choices without sufficient personal experience of the alternatives. This is achieved by providing information about the relative merits of alternative courses of action [RV97, SV99]. In contrast to traditional decision support systems, which are predominately used by specialists ( e.g. managerial decision makers), recommender systems are designed to support consumers in the decision making process [HN05, TAOl, SV99]. In the context of e-commerce applications recommender systems are used to *suggest products and services*  to users [Bur02, SKROl].

Recommender systems are also referred to as *recommendation systems.* In early publications (e.g. (GNOT92, RV97]) the term recommender system *was*  closely tied to a specific method of generating recommendations - namely collaborative filtering. This perspective limits recommender systems to a group of systems which uses a distinct method of generating recommendations (methodical view). Because of this narrower perspective the term recommendation system was proposed as a broader term, which denotes a system whose objective is to give recommendations regardless of the underlying recommendation

method (functional view) (SV99]. However, nowadays the tem1 recommender system is more frequently used in literature for both perspectives.

In this book the following *working definitions* are used:


The main objectives of recommender systems are to reduce information overload and improve decision quality (Run00]. Information overload occw-s because e-commerce stores may offer a wider range of products and services to the customer compared to traditional brick and mortar stores. In e-commerce stores the offered mix of products and services is not limited to physical space restrictions of the sales room. Thus, recommender systems are used to offer a subset of the product and service mix to the consumer to reduce information overload. Further objectives may be to provide personalized product information, rank items (i.e. products) according to the individual user profile, forecast user preferences for a distinct item, provide community critiques, and summarize community opinion (Run00, SKROI].

### **2.2 Classification**

Figure 2.1 shows a classification of recommender systems that considers four dimensions and gives an overview of the design alternatives of recommender systems:

Figure 2.1: Classification of recommender systems


teracting with the e-commerce application. *Pull technologies* require the customer to explicitly request recommendations, i.e. the communication process is initiated and controlled by the user [MGL97]. *Passive recommendation delivery* refers to presenting the recommendation in the natural context of the e-commerce application ( e.g. displaying recommendations during viewing or ordering a product). The advantage of this approach is to give recommendations when the user is already receptive to the idea of buying or consuming articles [SKROl].


A detailed description of these functional aspects of recommender systems based on this classification scheme can be found in Chapter 3.

Recommender systems may combine different methods of data acquisition, information delivery, and recommendation and vary the degree of personalization to best fit the user's needs [Bur02, BHC98, BS97). It may be useful to forego personalization in early stages of customer interaction. At this stage, data in the user profile is still sparse, trust in the e-vendor may be low and concerns regarding privacy may be high. Thus non-personalized recommendations based on statistical summarization or manual selection may be a good starting point to foster the relationship with the customer. After the successful establishment of a relationship and having overcome the initial barriers the e-vendor may add personalized recommendations to the customer interaction. For example amazon.com applies 18 different types of recommender systems with varying degrees of personalization, different methods of information delivery, diverse recommendation methods and varying input data on their web-site [GGSHST02].

Hence, an ideal recommender system should [AEK00]:


Degree of **personalization** 

### **2.3 Application Models of Recommender Systems**

As mentioned in Section 2.1 from a user's (i.e the customer's) perspective recommender systems reduce information overload, provide personalized product information, rank items, forecast user preferences, provide community critiques, and summarize community opinion. From the e- vendor point of view recommender systems ideally assist him or her in turning new and infrequent visitors of the web-site into buyers, building credibility through community inputs, inviting customers back, improving cross sales, and building long term relationships [SKROl]. Figure 2.2 shows five application models of recommender systems with their corresponding business goals. The degree of personalization - i.e. the extent of treating each customer individually - increases from the bottom to the top.

#### **2.3.1 Broad Recommendation Lists**

One of the most compelling challenges for e-commerce vendors is to turn *visitors into buyers.* Especially new and infrequent visitors need support in the navigational process to direct them to engaging products. E-commerce sites

use broad recommendation lists to give an overview of the range of products and services. The recommendations presented to the customer are not personalized and manual selection or statistical summarization are employed as preferred recommendation methods. These broad recommendation lists typically include overall best sellers, best sellers in a category, experts recommendations and other collections generated through manual selection or statistical summarization [SKROl].

Figure 2.3 shows an application of broad recommendation lists at Barnsandnoble.com. These broad recommendation lists are based on sales of Barnsandnoble.com and are updated hourly. Besides the overall best sellers in the category "books" this e--vendor offers best sellers lists in other product categories ( e.g. DVDs, videogames etc.) as well as best seller lists in different subcategories of books (e.g. adult fiction, business).

One major advantage of broad recommendation lists is the low degree of personalization. Thus the required amount of personal inf01mation about the user is low ( e.g. ephemeral context information about the category of interest to the user). This makes broad recommendations appropriate in early stages of customer interaction, when the customer is reluctant to give personal information to the e--vendor. Products suggested in broad recommendation lists are inherently appealing to the majority of the customers. Hence, they are not suitable for users interested in niche products. Without personalized recommendations it is indeed very difficult to meet the taste of these users.

#### **2.3.2 Customer Comments and Ratings**

Another business goal e-commerce vendors try to achieve with recommender systems is to *build credibility through community.* The e-commerce application should support the community of users as a platform for customer comments and ratings. This may help to overcome the problem of a possible initial distrust of the customer in the e--vendor. Usually customer comments and ratings are displayed in addition to the e--vendor's product descriptions. They function as a trust building measure, because the customers usually have more confidence in the opinion of other customers [SKROl].

Figure 2.3: Barnsandnoble.com overall best seller list in the category "books"

Figure 2.4 illustrates customer comments and ratings at the Amazon.com. The user may rank the product based on an ordinal scale from one to five. In addition to this purely quantitative rating a qualitative review in form of a textual description (limited to 1000 words) is also possible. These textual reviews are of major importance especially when personal taste is a significant criterion for the purchase of the product (e.g. books, music). Amazon.com uses mechanisms to ensure quality of the customer reviews by enabling other users to submit meta-recommendations for reviews. The reviews voted most useful by the Amazon.com community are displayed first ( "Spotlight reviews"). Further Amazon.com has set up several incentives (e.g. vouchers) to enhance community activities.

Figure 2.4: Customer comments and ratings on CDs at Amazon.com

An advantage of customer comments and ratings is that they require little effort by the e-vendor because all evaluation is done by the customers. However, the e-vendor must focus on usability of the e-commerce application to provide a comfortable platform for community communication and provision of advice or feedback on products. As a further benefit, community related initiatives may help to distinguish the e-vendor from competitors.

#### 2.3.3 Notification Services

Notification services are an application of push communication in recommender systems to invite existing customers back to the store and increase sales. Notifications are typically sent via e-mail when new products are in stock or special offers are available. A simple and often used form of notification services enables the customer to specify attributes ( e.g. category of music or book, price range) of products he or she is interested in. When the desired products are available, the user gets a message from the e-vendor. These services are a good starting point for individualized customer interaction.

FigW"e 2.5 shows an example of a simple notification service based on user pre-selections. Educanext.org is a platform for exchanging higher education learning resources. The user may subscribe to different academic disciplines he or she is interested in. When new learning resources in the specified fields are uploaded to the platform, the user receives an e-mail that lists titles and authors of these new resources.

However, more complex personalization techniques go beyond these simple pre-selections of attributes by the user. They monitor user behavior, build dynamic user profiles and adapt recommendations towards individual users based on the profiles.

#### **2.3.4 Product Associated Recommendations**

A further business goal for recommender systems is to *increase cross-sales* by means of product-associated recommendations. In brick-and-mortar stores complementary products are arranged nearby to encourage cross-sales. Since e-vendors do not have th.is spatial arrangement opportunity, recommender systems may suggest related products. Moreover recommender systems may go a step further and use the user profile to provide personalized cross-sales lists. A variety of input data may be used to generate such cross-sales lists. This includes anonyn10us purchase histories, customer purchase histories, ratings, product attributes, and expert opinions [SKROl]. Another option is to use explicit community knowledge to create or improve such lists.

As shown in Figure 2.6 Musicstore.de suggests complementary products ( "suitable accessories") based on specific product attributes. As a further example, Amazon.com employs past buying behavior of other users to create such cross-sales lists ( "Customers who bought this title also bought"). In addition

#### *CHAPTER 2. RECOMMENDER SYSTEMS* - *DEFINITION, CLASSIFICATION, AND MARKETING PERSPECTIVES*

Figure 2.5: Notification service of Educanext.org 

customers from Amazon.com may be explicitly recommended complementary items ( "Our customers' advice"). 

#### **2.3.5 Persistent Personalization**

One of the most challenging goals of recommender systems is to *build longterm relationships.* Long-term relationships should increase sales volume per customer and should help the e-vendor to build competitive barriers. This may be achieved by persistent personalization. Personalized recommender systems a.re based on the customer's history of preferences, purchases, or navigation and try to meet the needs of each individual customer. Personalized recom-

Figure 2.6: Cross-selling based on product attributes at Musicstore.de

mender systems dynamically learn user interests and store them in the user profile of the customer each time he or she interacts with the e-commerce application |SKR01].

Usually personalized recommender systems use information filtering techniques (e.g. user-based collaborative filtering) to address each customer individually. Persistent personalization raises competitive barriers, because by and by the e-vendor can meet the customers' needs more specifically and improve loyalty. The time consuming character of the learning relationship between e-vendor and customer hinders the customer to switch to another e-vendor easily (i.e. switching costs).

#### *CHAPTER 2. RECOMMENDER SYSTEMS* - *DEFINITION, CLASSIFICATION, AND MARKETING PERSPECTIVES*

Figure 2. 7: Persistent personalization based on collaborative filtering at Movielens.emu.edu

Figure 2. 7 illustrates personalized recommendations at Movielens.emu.edu by applying collaborative filtering in conjunction with explicit user input. Movie-Lens is a non-commercial research site run by GroupLens Research at the University of Minnesota. On this site the user explicitly rates movies he has already seen. This information is stored permanently in the user profile.

#### **2.4 The Consumer Decision Process**

As mentioned in Section 2.1 recommender systems assist the consumer in the decision making process. Hence, understanding this process may provide helpful insight when a vendor plans to apply a recommender system. In this section a holistic model of the consumer decision process as proposed by Blackwell et al. is presented (BMEOl]. In contrast to partial models of consumer behavior, holistic models try to integrate and interrelate all fundamental constructs of consumer behavior in regard to the decision process (MefO0]. Figure 2.8 shows the *phase model* of this process (BME0l], that includes seven phases: (1) need recognition, (2) search for information, (3) pre-purchase evaluation of alternatives, (4) purchase, (5) consumption, (6) post----consumption evaluation, and (7) divestment. This model represents a roadmap of consumers' minds, which is relevant with respect to recommendation applications of e-vendors. Consumers may be supported in the individual phases by recommender systems as described in the following sections.

#### **2.4.1 Need Recognition**

*Need recognition* occurs, when an individual senses a difference between what he or she perceives to be ideal in contrast to the actual state (BME0l]. As shown in Figure 2.9, need recognition appears, when a certain degree of discrepancy between the actual state (i.e. the consumer's cmrent situation) and the desired state (i.e. the situation a consumer wants to be in) appears. When a given level of threshold is reached, the consumer becomes aware that he or she has a need, that probably can be satisfied through a product or service.

Need recognition may either happen for reasons outside the control of a company or may be influenced by businesses. Advertising is a possible instrument for companies to generate needs [OM98). Especially personalized recommendations provided by recommender systems can be understood as a form of "advertising tailored towards the individual". Hence, recommender systems may be used to create or stimulate these needs more efficiently. In this stage of the consumer decision process push----communication may be a reasonable

#### *CHAPTER 2. RECOMMENDER SYSTEMS* - *DEFINITION, CLASSIFICATION, AND MARKETING PERSPECTIVES*

Figure 2.8: The consumer decision process(BME0l]

method to effectively make customers aware of their desire. The needs of customers a.re essentially influenced by two factors: (1) *environmental influences*  (e.g. culture, social class, personal influences) and (2) *individual differences*  (e.g. consumer resources, motivation, attitudes, knowledge) (BMEOl]. If recommendations take these two factors into consideration, they may effectively assist the consumer in identifying his needs.

#### **2.4.2 Information Search**

*Information search* is the next step in the consumer decision process model. Once a need is recognized, consumers starts to search for information to satisfy the unmet needs. This search may occur *internal or external.* Internal search refers to retrieving decision-relevant knowledge from memory. In contrast ex-

Figure 2.9: The need recognition process[BMEOl]

ternal search occurs when the consumer is collecting information from the marketplace, peers or other relevant information sources. Figure 2.10 shows the com1ection between internal and external search. External search usually occurs after the internal search process [Pun87]. ff the consumer thinks, that his knowledge is inadequate for the purchase decision he or she probably will undertake external search. This may happen passively (i.e. the consumer becomes more receptive to information sources) or actively, when the consumer exhibits search behavior like screening consumer publications, advertising material, web-sites or venturing retail outlets. External search can be categorized in *pre-purchase search* and *ongoing search.* Pre-purchase search is motivated by an upcoming purchase decision, whereas ongoing search is happening on a regular basis regardless of sporadic purchase needs [Pun87, BMEOl]. Recommender systems may be used to assist the consumer in both categories of external search. For instance, if a book enthusiast gets recommendations of new publications in his or her fields of interests sent by e-mail on a regularly basis, he or she is supported in the process of ongoing search.

When the consumer applies external search the following steps are involved to process information [BMEOl]:


External information sources can be categorized as ( 1) marketer-dominated and (2) non-marketer-dominated [BMEOl]. *Marketer-dominated* sources are provided by vendors for purposes of information and persuasion ( e.g. advertising, weo-sites, salespersons). However, *non-marketer dominated* sources like friends, families, opinion leaders and media may be even more influential to customers decisions than marketer-dominated information. By building virtual communities and employing recommender systems, vendors may utilize this kind of information to build credibility. For example, recommender systems may summarize community critique and recommend products with high ratings from the virtual community members or experts. By doing this, vendors may assist the consumer in the decision making process by providing nonmarketer----dominated information. However, in order to build or maintain credibility it is crucial to use this information sources honestly. For exan1 ple if it turns out that a vendor manipulates or censors community opinions wrongfully, severe implications in regard to the credibility of the vendor may occur. Thus, a vendor should publicize codes of conduct or ethical guidelines, how he or she deals with information provided by customers or third parties in general.

Figure 2.10: The internal information search process[BMEOl]

In this context the question arises, how extensive consumers conduct external search. The framework of *"economics of information''* as proposed by Stigler [Sti61] provides an insight to this problem from a cost-benefit perspective. According to this framework consumers inform themselves about products and services on the market to the point where the marginal costs of gathering more information equals or exceeds the marginal return (i.e. the benefits from gathering new information) [Urb86]. A study conducted by Srinivasan and Ratchford identified perceived risk (i.e. the consumers' uncertainty about the potential positive and negative consequences of the purchase decision), amount of experience with the product class, content of experience (i.e. positive or negative), and cost of search as essential determinants of the an1ount of search effort [SR91]. Because online recommender systems can reduce search costs significantly, they a.re a valuable tool for consumers with respect to external search.

#### **2.4.3 Pre-Purchase Evaluation of Alternatives**

In this stage of the consumer decision process the focus is on the manner in which consumer evaluate purchase alternatives [BMEOl). Before making a purchase decision, consumers usually compare and contrast different products and services. Consumers may use already existent or new evaluations stored in memory to select products and services that will most likely satisfy their needs. How this process is undertaken is again influenced by *individual differences* and *environmental influences.* In this process *salient* and *determinant* attributes are distinguished [BMEOl]. The consumers judge salient attributes as the most important characteristics of a product or service ( e.g. price, processor speed and size of the hard-disk of a personal computer). However, the conswner applies determinant attributes to actually select a certain product and service, especially when the salient attributes are considered as equal between the alternatives. Determinant attributes turn out to be often very subjective to the personal taste of the consumer (e.g. design of the personal computer).

Figure 2.11 shows the pre-purchase evaluation process. When a decision has to be taken, consumers usually do not consider all available options. In fact they limit the alternatives to a subset called the *"consideration set"* [RL91]. When consumers are evaluating alternatives the may (1) rely on pre-existing evaluations stored in memory (in this case the consideration set is called the "retrieval set") or (2) decide to construct new evaluations based on information acquired through internal or external search [BMEOl].

*Pre-existing evaluations* may be based on the consumers own past purchase and consumption experience with a product or service. In other cases - especially when the consumer has a lack of own experience - indirect or secondhand experiences ( e.g. impressions heard from friends) may become dominant for the evaluation. This illustrates the importance of word-of-mouth in the decision process. When consumers are unable (e.g. lack of pre-existing experience) or unwilling ( e.g. changes in environmental factors) to rely on pre-existing evaluations, they may decide to *construct new evaluations.* At this consun1ers may apply two basic processes: (1) the *categorization process* or (2) the *piecemeal process* [Suj85).

The *categorization process* refers to the evaluation of alternatives in respect of

Figure 2.11: The pre-purchase evaluation process[BMEOl]

particular mental categories to which they are assigned. The basic assumption is that people naturally divide the world of objects around them in categories, permitting an efficient way of processing and understanding of the environment [Suj85]. These categories may range from very general (e.g. computing machines) to very specific (e.g. laptop computers from Apple). Consumers typically assign their mental categories some degrees of liking or disliking. Furthermore, the evaluation attributed to a specific category may be transferred to any new object of that specific category [BMEOl]. On a regular basis, this is how consumers form evaluations of new products and services. To the extent that the new products or services a.re assigned membership to a given categmy, they will be evaluated with respect to the degree of liking of that category. This process of retrieving evaluations can also be referred to as a "schema-driven affect" , because typical category "exemplars" or "prototypes" function as a. scheme for the evaluation process (Suj85]. "Exemplars" are well-lmown actual examples of the category, whereas "prototypes" a.re abstract fictional images of the category, that embody typical attributes and characteristics associated with the category.

A more complex method to evaluate products and services is called the *piecemeal process* [BMEOl]. In this case, products are evaluated on a attribute- by-attribute basis. Products are perceived as a bundle of discrete attributes, with each attribute having a distinct subjective value or weight [Suj85]. The piecemeal process can be divided in three phases: (1) determination of important criteria or product dimensions, (2) judgement of the decision alternatives in view of each single attribute, (3) judgement of the overall performance of the alternatives.

In the first place, consumers must determine the *product dimensions* ( e.g. processor speed, memory size, price of a personal computer), they want to employ in the evaluation-process. Further important dimensions are the feelings that come from owning and using a certain product ( e.g. prestige, status, excitement). When decisions include "non-comparable" alternatives ( e.g. a consumer has to choose between different product categories) more abstract criteria have to be employed, because the alternatives share only a few common criteria along which comparisons can be undertaken [Joh89, BS87]. For instance if a consumer has to decide between different forms of entertainment ( e.g. buying a home stereo vs. buying a gaming console), more abstract criteria - like status or necessity - have to be used for comparisons.

The next step requires the consumer to evaluate each product and service in the consideration set along each criterion, that was judged as important before. As mentioned in Section 2.4.2, consumers perform internal (i.e. information already stored in memory) and external search to evaluate alternatives [SR91]. So called *"cutoffs"* are often used by consumers to simplify decision making [KB87]. A cutoff represents a predetermined acceptable level for an attribute. For instance, if a price of a product exceeds a certain acceptable limit, the product will be eliminated from the consideration set. *Signals*  are a further important component in evaluating product attributes. In general, signals are product attributes that consumers use to infer other product attributes (e.g. price as an indicator of high quality) [BMEOl, DMG91].

The third and final step in the piecemeal process is the judgement of the overall performance of the alternatives in the consideration set. Consequently, this is derived from the evaluation of the performance of the alternatives in respect of each attribute. Research literature has identified a number of ways

how consumers perform this task [EJW04, BME0l]. In principle *compensatory*  and *noncompensatory* evaluation strategies can be distinguished.

*Noncompensatory evaluation strategies* refer to an evaluation process, where a product's weakness on one attribute can not be compensated by its strong performance on other attributes [BME0l]. Noncompensatory strategies are applied in different forms [BME0l, EJW04, GW84]:


*Compensatory evaluation strategies* occw-, when the consumer accepts that poor ratings on some of the attributes may be offset by excellent ratings on

other attributes. Consequently a perceived weakness of an attribute ( even the most important one) may be compensated by other attributes. *Simple additive* and *weighted additive* a.re prominent forms of compensatory evaluations strategies [BMEOl, AM87]:


Understanding these strategies is an important issue when designing a *recommender system.* These systems a.re also in need of an "evaluation strategy" to determine, how much a consumer will like a certain product. The methods of generating recommendations may range from very simple ( e.g. nonpersonalized recommendations based on statistical summarization) to fairly complex (e.g. personalized recommendations). For a detailed description of these methods see Section 3.5. For instance, if personalized recommendations are generated by means of attribute-based filtering (see Section 3.5.2.4), evaluation strategies of consumers a.re closely related to the *classification algorithm*  (i.e. the algorithm to estimate the degree of interest in the product or service). If the chosen classification algorithm mimics the evaluation strategy of the consumer successfully and explains these assumption transparently (for explanations in recommender systems see Section 3.2), the consumer is likely to accept the recommendation.

#### **2.4.4 Purchase**

The next two stages in the consumer decision process model a.re purchase and consumption. Figure 2.12 summarizes, how the stages one to four (i.e. need

Figure 2.12: The consumer decisions process: Purchase[BMEOl]

recognition, search for information, pre-purchase evaluation, and purchase) of the consumer decision process model fit together:

In the purchase decision process consumer decide: (1) whether to buy, (2) when to buy, (3) what to buy, (4) where to buy, and (5) how to pay. At this pmchase decisions may occur in three different forms [BMEOl].

1. Fully planned purchase: A purchase is referred to as *fully planned,* when both the product and the brand are chosen in advance [BMEOl]. Consequently the consumer focuses his attention toward a specific product or service when interacting with the e-commerce application. In e- commerce applications recommender systems may be used as a marketing tactic to divert the consumers attention to other brands. For instance products with similar characteristics but better margins of profit may be presented to the consumer when he adds a product to the virtual shopping basket. However, consumer may perceive this as disturbing, if he thinks that this kind of recommendations simply favors the e--vendor. Consequently, these "substitutive" recommendations must also offer a benefit to the conswner ( e.g. suggesting a special offer with a better price/pe1formance ratio). A less intrusive option is to display complementary products to increase cross-sales. For instance, if the conswner buys a specific digital audio player, a docking station and a protective cover may be recommended to him.


Figure 2.13: A product finder for digital cameras

Recommender systems may also be used to support this *impulse buying behavior.* For instance if a consumer adds a CD of a certain artist to his virtual shopping basket, buying a printed biography of that artist may be suggested to him or her. Studies show that unplanned purchases play a major role in "real world shopping" trips. Consequently, recommender systems may function as a vehicle to gain extra revenues in e-commerce applications by supporting impulse buying behavior [SKROl].

#### **2.4.5 Post-Purchase Processes**

The post-consumption processes include consumption, post-consumption evaluation and divestment. *Consumption* refers to the usage of the purchased product and service. *Post-consumption evaluation* is a further fundamental part of the consunier decision process model. During and past the consumption consumers form evaluations in regard to the product and the consumption experience [BMEOl]. *Divestment* constitutes the final stage of the model. At this consmners may resell, dispose or recycle the product [BMEOl].

#### *CHAPTER 2. RECOMMENDER SYSTEMS* - *DEFINITION, CLASSIFICATION, AND MARKETING PERSPECTIVES*

Figme 2.14: Three types of consumption experiences (BMEOl]

*Consumption* is always connected to experiences, that can be categorized as (1) positive reinforcement, (2) negative reinforcement, and (3) punishment [BMEOl]. Figure 2.14 gives an overview about these three types of consumption experiences.

*Positive reinforcement* occurs when the consumer receives a positive outcome from the product usage. For instance, playing a thrilling video game or reading an interesting book is regularly connected to positive reinforcement. *Negative rein/ orcement* emerges, when the consumption of a product or service enables the conswner to avoid or minimize negative outcomes. Vaccination is a typical example for negative reinforcement, because it prevents from getting sick. The third type of consumption experience is referred to as *punishment.* Punishment happens when the consumer receives negative outcomes from the product usage (e.g. listening to a CD the consumer dislikes). If punishment is experienced, it is quite unlikely that repeat usage or repeat pm-chase will occm· (BMEOl].

The *confirmation or disconfirmation of expectations* that carried the consumer into purchase and consumption is of further interest. These expectations have a massive influence on the post-consumption evaluation.

*Post-consumption evaluations* a.re formed during and after consumption. Postconsumption evaluations may resemble pre-purchase evaluations, especially when the consumer is satisfied with the product or service. In other cases, post-conswnption evaluations may differ substantially from pre-purchase evaluations [BMEOl]. In this case, the product may either do not meet the user expectations or pedorm significantly better than expected ( which is the less frequent case, because low pre-purchase expectations seldom result in purchases). Post consumption evaluations are of great importance for companies. They (1) influence repeat buying behavior, (2) shape word-of-mouth communication, and (3) lead to complaints due to dissatisfaction.

*Repeat buying behavior* usually emerges, when the consumer is satisfied with products or services. Hence, positive post-consumption evaluations are crucial for *retaining c'UStomers* [BMEOl]. This is of major importance for companies, because it is much cheaper to retain old customers than to gain new ones [FW87]. Consequently, marketing concepts like relationship marketing or one-to-one marketing have emerged. This concepts put heavy emphasis on customer retention. Recommender systems may further contribute to the retention of customers in respect of e---vendors. If recommendations are perceived as useful, they represent a *value-added service,* that leads to higher customer satisfaction. Especially personalized recommendations that are based on a long-term learning relationship foster the relationship between customer and company. In this case, switching costs arise for the customer. These switching costs hinder the customer from easily moving to another e---vendor.

*Word-of-mouth communication* is a further consequence of post-consumption evaluations. It is a common activity, that consumers are discussing their consumption experiences with others. Usually word-of-mouth communication resembles the outcome of post-consumption evaluation. Hence, the favorability of word-of-mouth communications is directly linked to the favora.bility of the consumption experience [NG05, Ric83]. A company's ability to provide a satisfying consumption experience will affect its ability in retaining current customers as well as acquiring new ones [BMEOl ]. In e-commerce applications,

word--of-mouth communication could be used for purposes of the vendor. In this connection, recommender systems in conjwiction with virtual communities offer a way to use word--of-mouth communication for recommendation purposes, for building credibility and to distinguish the vendor from others. For instance, if recommendations a.re given, *customer comments and ratings*  may be displayed to enrich the vendor's product or service description. This helps to build trust in the e---vendor. Further, some recommendation methods ( e.g. collaborative filtering) require ratings of customers to generate recommendations. If a vendor wants to employ these recommendation methods a lively virtual community is a must. Additionally, customer comments and ratings may assist to improve the mix of products and services offered to the customer by eliminating products that cause massive dissatisfaction. Clearly, dissatisfaction is also reflected by decreasing sales volwnes in the long run. However, using customer comments and ratings enables the vendor to react faster to dissatisfaction. For manufacturers and service providers customer comments and ratings are a valuable source of information for product or service improvements.

*Complaints* are a further consequence of dissatisfied customers. Companies should encourage customers to communicate complaints. Corrective actions to avoid or minimize future unhappiness can only be ta.ken, if the company knows the reasons for dissatisfaction [BME0l]. Hence, companies should make it *as* easy *as* possible for customers to file their complaints. A sincere and quick response to complaints may alleviate dissatisfaction and may even lead to stronger repurchase intentions [Gil82, BMEOl]. Additionally, enabling the customer to express his dissatisfaction leads to significantly less negative wordof-mouth [NG05]. As a consequence, e---vendors should support the submission and management of complaints in their e-commerce application.

### 2.5 **Virtual Communities**

Virtual communities are an important factor in e-commerce applications and recommender systems respectively. In general virtual communities are *social networks that use computer-mediated spaces* ( e.g. the Internet) for communication [HA97, LVL03, And02, Koz99]. They offer a potential for an integration

of content and communication with an emphasis on member-generated content [HA97]. In virtual communities people (e.g. consumers) interact socially for mutual benefits (And02].

Virtual communities may be classified along the desire to meet four basic needs: (1) interest, (2) relationship, (3) fantasy, and (4) transaction [HA97]. *Virtual communities of interest* bring together people that share an interest and an expertise in a specific topic (e.g. music-lovers). *Virtual communities of relationship* consist of people who have similar experiences. The community enables them to come together and form meaningful relationships ( e.g. people with a certain disease). *Virtual communities of fantasy* give people the opportunity to come together for entertainment purposes (e.g. role-playing gamers). *Virtual communities of transaction* have the purpose to connect people, who want to trade information, products and services ( e.g. communities located at eBay or Amazon).

#### **2.5.1 Characteristics and Benefits**

For the scope of this book, *virtual communities of transaction controlled by evendors* are of special interest. In general, these virtual communities may be operated by vendors or manufacturers (i.e. "seller controlled") or independent third parties (i.e. "neutral"). In B2B-environments communities of transaction may additionally be controlled by buyers.

*Communities of transaction* controlled by e-vendors share the following characteristics (HA97, SG00]:


is to support the customer in the buying decision process by providing additional member-generated content.


*Virtual community of transactions* offers the following benefits to the operator (i.e. thee-vendor) [PR98, HA97):


Figure 2.15: Reinforcing virtuous cycles in e-vendor controlled virtual communities of transaction (adapted from Paul and Runte [PR98])

#### **2.5.2 Virtual Communities and Network Effects**

Virtual commmtlties are subject to *positive network effects* [Lie02, HA97]. A positive network effect means that the value of a virtual community grows with the number of its members. That circun1Stance may ultimately result in increasing returns for the operator of the community [HA97, Art96]. This is caused by a series of interacting and reinforcing virtuous cycles shown in Figw·e 2.15 [HA97, PR98].

As Figure 2.15 illustrates, the reinforcing virtuous loops refer to [HA97, PR98]:

• Member-generated content: The basic asswnption is, that membergenerated content is a key source of content attractiveness. That content instigates members to join and remain in a virtual community. As a consequence, the more members a community has, the more content is created. This in turn raises the attractiveness of the community, which causes more people to join the community.


As a result from these self-reinforcing cycles, managing member evolution is a key success factor of virtual communities (HA97, PR98, AndOl]. When a critical mass of members is reached, network effects lead to a self-reinforcing growth of contents, member profiles, loyalty and transactions. In the following chapter problems related to the successful building of a virtual community is discussed.

#### **2.5.3 Community Building**

One of the most challenging problems of setting up a virtual community is to achieve a critical mass of members. Hence, e-vendors should asses their potential to control a commwtlty carefully [And0l, HA97). The potential of a successful community depends on ( 1) *indicators of the economic potential* and (2) *resources* of the community organizer [HA97).

*Indicators of the economic potential* include (1) the size of the potential community, (2) the relative value of being online, (3) the value of being in a community, ( 4) the likely intensity of e-commerce, ( 5) the fractal depth of the corrnnunity, and (6) the fractal width of the community [HA97).

Estimating the *potential size of the virtual community* can be done by referring to demographic statics. For instance, a book-seller may focus on a specific area (e.g. German-speaking countries). Another factor that is of interest is the spending information of the individual consumers. Spending information helps to assess the overall market size in terms of money and potential sales volume for the e-vendor. A fmther determinant of the potential size of the community is the number of people buying information about the specific field of interest. For instance, how many people do subscribe to music-related journals or magazines? Answering this questions may help to determine the relevancy of a virtual community for these people. Another factor that may help to estimate the size of the corrnnunity can be membership in associations or groups. This factor clearly shows the importance of social networks in the relevant field [HA97).

Firstly, the *relative value of being online* refers to the number of people, who have to ability to join a virtual community because of they are physically equipped to go online. For instance, a virtual community for well-educated and ntlddle--aged people is more likely to be successful compared to communities who aim at elderly and poor people. The second aspect is the relative value of the online--corrnnunity compared to off-line alternatives. If the virtual community is cheaper, more efficient and offers unique capabilities it is likely to prosper. For instance many online newspapers or magazines add the ability to comment articles by community members. This creates an value-added service because people are often interested in the opinion of others. This service offers a chance to discuss with like-minded persons and to f01m social networks with them. Further, the relative value of virtual communities that focus on markets that are fragmented or where geography creates barriers is regularly very high IHA97J. For instance, communities of transaction that focus on spare parts for rare old-timer cars may be successful because of this.

The *value of being in a virtual community* refers to the intensity of satisfying needs IHA97]. In community of transactions these needs are usually related to the products and services the community focuses at. If the products are complex, hard to evaluate and complicated to use ( e.g. sophisticated software) it is very likely that the virtual community assists the members in solving product-related problems. Here, experiences of other purchasers of the same goods constitute a valuable source of information.

Especially in virtual communities of transaction the *likely intensity of ecommerce* is of major interest. The operator of such a community must estimate the overall volume of transactions conducted by the targeted community group and the average size of each transaction IHA97]. In this context, characteristics of the products and services ( e.g. size and bulk relative to value, thin markets, perishability, immediate gratification factor) offered by the e--vendor are of major importance. For a discussion of products that are likely to create a large transaction volume by e-commerce applications see: [HN05, Lie02]

The *fractal depth of the community* is the degree to which it can be segmented into sub-communities. The more ways a community can be split, the more it can create small and focused sub-communities. In these sub-communities the participants are more likely to have common interests. As a consequence, the members will be more dedicated to the sub-commwuty and spend more time online. Fmther, members are more likely to engage in transactions [HA97]. For instance, a travel community can be split by regions, by travel type ( e.g. air travel, train jow·neys), and by reasons for travel.

*Fractal breadth of the community* refers to the ability of the community to build out to arenas that bear no relation to the community's original focus IHA97]. This may enable the e--vendor to extend the offered range of products and services. For instance, a book-seller with a lively community may have an advantage, if the vendor decides to offer CDs additionally. It is likely that

synergy effects will occur, because community members will also engage in rating CDs and making comments on them.

Besides the indicators of the economic potential mentioned above, the following *resources* of the e-vendor ease the building of a community especially in the early stages: {1) brands, (2) customer relationships, and {3) content.

A *strong bmnd* carried over to the online world is a valuable asset for attracting customers to a web-site. Brands help to establish trust and credibility especially in the early stages of the community. Hence, brands make it easier to reach a critical mass of community members and to set the reinforcing virtuous cycles into motion [HA97].

*Established customer relationships* are a further benefit in the early stages of community building. Customer relationships can be understood as strong understanding of what the individual customer wants and an ability to deliver what the customer needs. They also imply an ongoing interaction with customers that constitutes an opportunity to introduce them to a newly established virtual community [HA97]. Regarding the ongoing interaction necessary for customer relationships, virtual communities may also help to reduce transaction costs for both the e-vendor and the customer since online communication is regularly cheaper.

*Published content* is a further key factor in the early stages of virtual communities. Since the volume of member-generated content is low in these stages, providing an interesting content is helpful to attract members, particulary if the content is adapted to make use of the special capabilities of the online medium [HA97]. For instance, a book-seller may buy in book-reviews from external sources. These reviews from experts may spm· community member to post their own opinions in the virtual community.

In the context of community building the typical member development path is of special interest. Figure 2.16 exhibits the four stages of member development.

The first step is to *attract members.* Marketing initiatives, attractive content, and free membership and usage are levers to allure new members to *CHAPTER 2. RECOMMENDER SYSTEMS* - *DEFINITION, CLASSIFICATION, AND MARKETING PERSPECTIVES* 

the community. The next stage is to *promote participation* in the community. For instance, community organizers should provide incentives to engage community members in providing member-generated content. In the following *building loyalty* is of central importance for community operators. Loyalty emerges by supporting member-to-member relationships and by fostering member-to-host relationships. For instance, existing customer retention strategies should be incorporated and adapted to the enhanced possibilities of the online medium. Finally, the e-vendor should *capture value* from the community engagement. Recommendations based on member-generated content (i.e. collaborative filtering, product-association rules, statistical summarization of community opinions) may be applied to increase transaction opportunities. A fmther possibility to capture value is to offer individualized products and services based on the information stored in the member profile [SG00].

For community building purposes it is of importance to understand that not all community members are equal in terms of their economic potential to the community [HA97, Koz99]. Figure 2.17 presents a classification of different types of members in communities of transaction [Koz99].

The formation of lasting identification as a member of a virtual community is largely determined by two factors: (1) the self-centrality of the consumption activity and {2) the intensity of the social relationships the person possesses with other members of the virtual community [Koz99]. The concept of *selfcentrality of the consumption activity* refers to the importance of the symbols of the particular consumption in respect of the self-image of a person [Koz99). For example, for book-aficionados reading books is a central activity to their psychological self-concept. The higher the self-centrality of the consun1ption activity the more likely a person will be to pursue and value membership in a virtual community. The second factor, *social ties to the community* is very often related to the self-centrality of the consumption experience. For

Figure 2.17: Types of virtual community members [Koz99]

instance, a young male who is extremely devoted to classic Italian scooters and who lives in a rural environment is likely to seek like-minded people on the Internet, especially if he has few people in his face-to-face community that share his passion.

As shown in Figure 2.17 *tourists* la.ck strong social ties and their interest in the consumption activity is only superficial or passing. Consequently, the interest in the products and services offered is very limited. *Minglers* a.re persons that maintain strong social ties, but show no deeper interest in the central consumption activity. In contrast devotees maintain a strong interest in and enthusiasm for the consumption activity. However, their ties to the virtual communities a.re low. The last category is called the *insider.* Insiders show strong interest in the conswnption activity and have strong personal ties to the community [Koz99].

From a marketing perspective *devotees and insiders* a.re the most important target group for communities of transaction. Because of their high self-centrality of consumption, these two types usually are "heavy users" of the offered products and services of thee-vendor. Thus, they will have a large share of transactions and sales volume respectively, especially when repeat-purchases are characteristic for the offered product category (e.g. books, CDs). Additionally, devotees and insiders regularly have a massive knowledge of consumption. This makes them a primary target for the contribution of member-generated content. In this context personalized recommendations are a good initiative to tie devotees to the community and convert them to "loyal" insiders, because of switching costs.

To get a better understanding of the interests of the different types of community members, different *social intemctions modes* are presented in Figure 2.18. As a consequence, community organizers may apply *intemction-based segmentation* for the separate groups. This will allow community organizers to better formulate strategies that recognize the differential opportunities and needs of devotees, insiders, minglers and tourists [Koz99].

As shown in the figure, the modes of interaction are classified along two criteria: (1) objective of communication and {2) orientation of the communication. The objective of communication may be autotelic or instrumental. Autotelic communication takes place for the sake of its own (i.e. it has an end in itself), whereas instrumental communication is used to as a means for the accomplishment of other ends [Koz99].

In general *devotees and tourists* are uninterested in building on.line social ties. In virtual communities these member-types tend to use the *informational mode*  of interaction. They primarily use online communication as a means for the accomplishment of specific goals ( e.g. improve the quality of their purchase decision by reading comments on products and services of other community members). The social orientation of their communication is individualistic. These two groups usually communicate in order to receive a short-term personal gain. In general they are using other community members resources and do not intend to returning anything of benefit to other individuals or the group as a whole (Koz99].

In the context of *recommendation applications* devotees and tourists try to benefit from recommendations. In general they are not prepared to make an effort by themselves by rating or commenting products and services (i.e

# **Orientation of**

Figure 2.18: Interaction modes in virtual comn1Wl.ities of transaction [Koz99]

"free-riding"). Hence, explicit methods of data acquirement are not suitable for these groups (for a detailed description of methods of data acquirement see Section 3.1). However, they may be a valuable source of information, if implicit methods are used ( e.g. click-stream analysis). Additionally, e-vendors should encourage devotees to share their knowledge of products and services by applying marketing initiatives (e.g. incentive programs). Because devotees and tourists pursue short-term goals, personalization strategies may not be applicable. Thence, non-personalized recommendations should be applied.

*Minglers and insiders* a.re usually far more social in their group commllll.ication behavior. As a consequence they often use the *relational interaction mode.* To them, the social contact in the virtual commwlity has a value in its own. Their focus is on long-term personal gain through cooperation with other commllll.ity members or the delineation and enforcement of communal standards [Koz99]. This makes this interaction mode the most valuable for recommendation applications. Clearly, insiders and minglers a.re a valuable source for member-generated content. Especially insiders a.re the primary target group for the provision of ratings and comments, because of their usually high level of product-related knowledge.

The *recreational mode* refers to interactions that are conducted for primarily selfish or short-term satisfaction. In this mode online communication itself is the goal. It mainly occurs, when synchronous communication is possible in the virtual community {e.g. chat rooms). A good example is the often insipid small talk in chat rooms. This form of interaction is mainly used by *tourists and minglers* (Koz99].

*Transformational interaction* occurs when community members strive for positive change in regard to their interests. It is focused on longer-term social gain. This mode of interaction is primarily used by *insiders and devotees* (Koz99]. The goals connected with this interaction mode may sometimes be antipodal to the interests of the e-vendor { e.g. empowerment of consumers, change in consumption behavior).

## **Chapter 3**

## **Recommender Systems** - **Functional Perspectives**

This chapter gives an overlook of the functional aspects of recommender systems. It deals with functional input and output of recommender systems, measurement scales for preference elicitation, information delivery aspects, and recommendation methods.

Input data of recommender systems are described *in* Section 3.1. Input data can be classified along the dimensions *duration, acquisition, originator and origin.* 

Section 3.2 deals with output data of recommender systems. Besides the recommendations itself, recommender systems may display *predictions, text comments and ratings* to the user. Further, possible approaches to provide supplementary *explanations* (Le why certain products are recommended) are presented. Finally, the basic flow of input and output data in e-commerce recommendation applications is illustrated.

Section 3.3 examines different *statistical measurement scales.* It focuses on metric scales for the elicitation of user preferences.

Section 3.4 refers to the information delivery of recommender systems. *Push, pull and passive* technologies may be used to suggestions, ratings and predic-

tions to the user. Push, pull and passive technologies refer to the extent of the user's initiative to get recommendations.

The chapter concludes with the introduction of different recommendation methods. *Personalized and non-personalized recommendation methods* and their corresponding advantages and disadvantages are described in detail.

### **3.1 Input Data of Recommender Systems**

This section deals with the functional input data of recommender systems. Recommender systems use input data to generate output in f01m of suggestions, predictions and ratings. Figure 3.1 illustrates a classification scheme for input data of recommender systems.

Figw-e 3.1: Classification of input data

Depending on duration of the user data storage, persistent data, ephemeral data, or a combination of both may be used for personalized recommenda, tions [MT02). *Ephemeral data* is used on a per session basis only and is deleted afterwards, whereas persistent input data is stored over different user interaction sessions. Thus, ephemeral personalization can be applied to users, who are not authenticated to the e-commerce application. It may be useful when users are new or are reluctant to give personal information to the e--vendor. For instance, the current navigation of an unregistered (i.e anonymous) user could be used to push recommendations based on that context. *Persistent data* is acquired over different sessions and stored in user profiles permanently. Thus persistent data storage allows improving the user-profile over time and collecting long-term preferences of the users of the e-commerce application.

Acquisition denotes how the input data is gathered from user interaction. *Explicit data* is intentionally submitted by the user to inform the recommender systems about his preferences (e.g. rating items on a nominal scale), whereas *implicit data* stems from monitoring user behavior ( e.g. browsing the product catalogue) [SKROl]. In this context data acquisition is related to user awareness. This denotes the extent to which the user is required to give inputs to the recommender system intentionally. Consequently user awareness refers to the user's state of mind while interacting with the e-commerce application.

The advantage of explicit approaches is that the users know their interest best and are in control of the recommendation process. However, explicit approaches put the effort of adapting recommendations towards the users. Further, the users have to learn to handle the input forms of the recommender system. Thence, complexity is increased from the users' point of view. Consequently the user-interfaces of recommender systems, which are operated by non-specialists per definition, have to be designed carefully in respect of usability.

The pros of implicit approaches are that no or little effort is put towards the users and that no special knowledge of the user is required. But the user loses control over the recommendation process. Further implicit approaches reduce transparency of recommendations, i.e. the user does not understand how recommendations are generated. Thus it is difficult for the user to develop a coherent cognitive model of the recommender system [Wae04].

*User interrogation* is the most commonly used explicit data acquisition approach. The user is required to fill out forms to describe interests or other relevant parameters (e.g. keywords and attributes of items). User interrogation is often applied to obtain ratings of items the user has already knowledge

of. These ratings may be based on an ordinal scale (e.g. "rate this item on a scale from one to five") or on a binary scale ( e.g. "do you like this item - yes or no").

*Recording user behavior* is an typical implicit approach. It does not require the user to intentionally engage in the data acquisition process. A simple approach would be to give recommendations based on the item the user is currently viewing. In ~ommerce applications the articles in the virtual shopping-basket, the articles bought in the past or other clickstream-data can be utilized for recommendation purposes. According to studies from Morita and Shinoda as well *as* Konstan et al. the time a user spends viewing an artefact is a appropriate indicator for the relevance to the user (MS94, KMM+97]. Hence, time spent to view articles can be used as implicit input data for recommendations, although this data may be biased (e.g. the user is interrupted).

Explicit and implicit approaches may be combined. Usually these combined methods use explicit approaches to gain knowledge about the user in the initial phase of the system use and change over to explicit approaches in later phases. For instance reference items can be used to create an initial *item space* ( also referred to as *document space* because this method was first applied on textual documents [FD92]). The user has to judge the relevancy of these reference items by explicit user interrogation. New items are compared to these reference items and are recommended if the similarity to these reference items, which were rated *as* relevant, exceeds a certain threshold. The advantage of this method is, that the effort of user interrogations is limited to the initial system use. However, from the user's point of view it is hard to estimate the usefulness in the beginning of the system use. Hence, he or she might not be willing to put effort into judging reference items, when he or she has little knowledge a.bout the advantages of using the system. Additionally ongoing bias of the users interests may occur, if certain areas of interests a.re not covered by the initial item space [HSSOl]. *Stereotypic inference* is another combined approach. A Stereotype is a collection of attributes that often cooccur in people, i.e stereotypes a.re typical characteristics of user groups in a given domain [Ric89]. Users a.re **asked** to provide personal information by explicit approaches in the initial phase of system use. These data is used to relate the user to a specific stereotype (i.e. default initial profile) (HSSOl].

This method helps to address the bootstrapping problem (i.e. giving suitable recommendations to new users). Consequently, stereotypes enable the recommender system to make plausible inferences on the basis of a substantially smaller number of observations of the user's behavior. Over time observations are added to the profile, which may enhance or override default assumptions about the user.

A further criterion to classify input data of recommender systems is the originator of the input data, whereby active user input, community input and input from others (e.g. editors, critics) can be distinguished.

*Act-ive user input* refers to the data generated through interactions with the active user (i.e. the user who currently gets recommendations). Active user data typically include:


*Community inputs* usually refer to the sum of all active user inputs. Besides those internal data (see below) community inputs may also include external data (e.g. item popularity in f01m of national best-seller lists). Generally spoken, community inputs comprise of data, which denotes how multiple individuals in the community or the commwli.ty *as* a whole perceive attributes of items ( e.g. book categories or film genres are derived from the consensus of the broader society) (SKR0lj.

*Text comments* are community inputs in form of textual descriptions of users' experiences with single products or services. Text comments may be very useful to enhance the decision making process of the active user. However, the user's effort of processing text comments fairly high, since the user must read this textual information and interpret to what degree these comments contain positive and negative attitudes toward the item.

To ease this procedure, textual comments are often supplemented by *scores or ratings* of users, which indicate the overall satisfaction with the item. Additionally, these individual ratings can be summarized ( e.g. by calculating the arithmetic mean) to get an quick overview of the users' average opinion.

Finally, depending on the source, input data can be classified into internal and external data. *External data* stem from third parties and may relate to items or users. For instance item-specific external data may be derived from third party electronic product catalogues with categorizations and descriptions of product attributes (e.g. genre and keyword classifications of books or films) (SKROlj. External item popularity (e.g. national best-seller list) is a further example for item-specific external data used for recommendation purposes. Typical user-specific external data stem from market research companies ( e.g. general demographic data of online-shoppers) and may also be used in the recommendation process. In contrast to external data *internal data* is exclusive to the e-commerce vendor. Thus, it is of major importance in regard to competitive advantages. Internal data is often generated automatically by the user's interaction with the e-commerce data (e.g. clickstream-data), but may also be rendered manually (e.g. broad recommendation lists based on editors' manual selections).

### **3.2 Output Data of Recommender Systems**

The outputs of recommender systems are suggestions of items (i.e. products and services). Additionally, the may display ratings, text comments, predictions and explanations.

*Suggestions* make the user of recommendations systems aware of items that

the e--vendor considers *as* useful to the customer. Phrases like "we recommend ... ", "try this". Other phrases ( "additional products", "supreme products") are used to indicate the cross- and up-sell potential of certain items. Recommender systems may suggest either only one item or may display multiple items to the user. When a set of items is recommended by lists, the order of items may be arbitrary, which means that the sequence of items does not reflect any order of preference for the user ( e.g. alphabetical). In the other case, the order of items may indicate predictions of the degree of interest to the user (i.e. the first item on the list is the best-fit recommendation).

*Predictions* are estimates of ratings, the user would give to items. They quantify, how much a user will probably like the recommended item and hence indicate the strength of an recommendation. Predictions may be personalized, which means that they are based on the stored preferences in an individual user-profile. Non-personalized predictions refer to estimates for typical community members [SKROl].

*Text comments and ratings* constitute further possible output data of recommender systems. Suggestions of items may be supplemented by text comments. Because text comments are not completely machine--W1derstandable, many e- vendors require the user to give an additional numerical rating to indicate the direction of the comment (i.e. pro or against the item). Especially, when the size of the community is large and the number of text reviews is high, the recommender system has to display a selection of text comments, because showing all text comments would lead to information overload. The selection of text comments bears the risk of biasing information (e.g. only positive comments are shown to the user). Hence, accompanying numerical ratings can be used to show an representative selection of comments to the user by choosing a proportional number of positive, neutral and negative comments. Another notion to address the problem of selecting text comments is to apply "meta-ratings" . Meta-ratings are ratings about ratings (respectively text comments). This means, that the usefulness and quality of text comments from the community are judged by the community (e.g. "Was this review helpful to you?"). In this case, the most appreciated text comments are displayed first.

In recommender systems *explanations* can be used to expose the reasoning behind an recommendation. They enhance transparency in the recommenda-

tion process. Thus, they may raise the user's trust in the recommendation process and may also improve the decision-making performance. The benefits of adding explanation capabilities to recommender systems a.re [Her99]:


Since recommendation methods range from relative simple to highly complex with large a.mounts of data and extensive computation (see Section 3.5), the provision of explanations may also vary in terms of complexity. Three possible models for explanations are applicable [Her99]:

• Data-explorative model: When this model is applied, the application lets the user explore the data on which recommendations are based. Mathematical processes behind the recommendations are not explained (e.g. because they are to complex for the "average" user). Because some recommendation methods use large a.mounts of data, initially only a selection of key-data are displayed to the user. Key-data are of significant relevance for the recommendation process. However, the user can navigate to other parts of the data. The data-explorative model allows the user to validate the recommendation by their own personal approaches. For instance amazon.com applies this model. The user may click on a link labeled "Why was I recommended this?" to see the relevant items for the recommendation process as shown in figure 3.2.


Figure 3.2: Explanations using the data-explorative model


#### CHAPTER 3. RECOMMENDER SYSTEMS - FUNCTIONAL PERSPECTIVES

Figure 3.3: Flow of input and output data

Figure 3.3 summarizes the basic flow of input and output data in e-commerce recommendation applications and denotes the software components of a typical recommender system.

As illustrated in the figure, an e-store component is in charge of the information delivery to the user by applying push or pull technologies. The component also forwards the interaction data to the user model builder, which constructs a long-term and/or short-term user profile and stores the user profile(s) in a database.

The user profile stored in the database is employed by the recommender component. This component generates the suggestions, predictions, and explanations and summarizes ratings by applying recommendation methods (e.g. collaborative filtering, attribute-based filtering). This process is typically based on session data, the long-term user profiles and item data. Additionally external data regarding users and items (i.e. data that stems from third party sources) may be used in this process.

### **3.3 Measurement Scales for Preference Elicitation**

Preferences of users are the most important data for recommendation systems. They are generally used as input data but - as mentioned in Section 3.2 - may also be displayed as an output in form of predictions. To measure or indicate these preferences, different *statistical measurement scales* can be applied. Measurement scales can be categorized into *nonmetric (qualitative)* and *metric (quantitative)* scales (HATB98, BEPW03].

Nonmetric scales include nominal scales, binary scales and ordinal scales. *Nominal scales* a.re classifications of qualitative attributes, characteristics or properties (e.g gender, color). *Binary scales* a.re a sub-type of nominal scales with exact two possible occurrences of an attribute ( e.g. yes or no, male or female, zero or one). Nominal and binary scales a.re the scales with the lowest level of measurement precision. Arithmetical operations can not be applied to nominal and binary scales, but it is possible to calculate the absolute and relative frequency of an attribute. With *ordinal scales* variables can be ordered or ranked, i.e. attributes can be compared by "greater than" or "less than" relationships. The ranking of variables is relative. However, it is not possible to determine the distance between two occurrences of a variable. Similar to nominal scales it is not possible to use any arithmetic operation. However, additional to absolute and relative frequency, quantile and median can be calculated.

Interval scales and ratio scales a.re both metric scales, which refer to quantitative measurable attributes (e.g. a.mount of time, size of an object, temperature). Metric scales have constant units of measurement, i.e. the distances between two adjacent points a.re equal on any part of the scale (HATB98]. *Interval scales* have arbitrary zero points ( e.g. temperature in Fahrenheit or Celsius). Possible arithmetical operations for transformations of the scale are addition or subtraction. Feasible statistical operations are (amongst others) to calculate the mean value and standard deviation. Interval scales are widely used for measuring preferences explicitly ( e.g. "rate this item on a scale from one to five"). These scores and ratings are regularly assumed to be based on interval scales. However, strictly speaking, ratings rest upon ordinal scales, because it can not be assumed, that equal distances between two adjacent points on the scale are given on any part of the scale. **In** spite of this, ratings are predominantly treated as interval scales ( e.g. building the mean value of all user ratings) [BEPW03]. In contrast to interval scales, *ratio scales* have an absolute zero point (e.g. weight, length, speed). They represent the highest form of measurement precision and all arithmetical operations are allowed. In the context of recommender systems, ratio scales are preferably used when preferences are surveyed by means of implicit data acquisition methods ( e.g. time spend viewing an item).

### **3.4 Information Delivery**

The output of recommender systems (i.e. suggestions, ratings, text comments and predictions) may be transferred to the user by push, pull and passive *information delivery techniques.* 

*Pu.sh technologies* refer to methods, where the suggestions are given to the user without requiring the users' initiative, i.e. the recommender systems initiates the communication process [MGL97]. A distinctive example for push communication is the use of e-mails to send recommendations to users on a regular basis (e.g. fixed time schedule). This has the advantage of giving recommendations to users without requiring them to interact with the e-commerce application. They can be understood as an promotional activity to invite users to return to the e-commerce vendor. However, if the user is not satisfied with the recommendations (e.g. due to lack of personalization) he or she might consider the e-mails mentioned in the example above as spam.

*Passive technologies* denote information delivery, which supplements the presentation of recommendations to the normal use (i.e. "the natural context") of the e-commerce application (SKROl]. Hence, the might be understood as a sub-class of push technologies. For instance, recommendations are displayed based on the item the user is currently browsing. Another exan1ple for passive delivery is the presentation of supplemental goods or special shipping options during the ordering process. At this time the user may be very receptive to the vendors idea of up- and cross-selling. A possible disadvantage of passive recommendations is that the user might not recognize them as recommendations [SKROl].

In contrast *pull technologies* require the user to take initiative to get recommendations. In e-commerce applications these is usually achieved by clicking on a link ( e.g. "yom recommendations"). Pull technologies are usually perceived as unobtrusive, because no recommendations are displayed unless the user wants them to see.

#### **3.5 Recommendation Methods**

This Section focuses on specific recommendation methods. Recommendation methods can be classified according to the degree of personalization. Methods for non-personalized recommendations do not refer to individual user profiles. Thus they give identical recommendations to different users. Methods for personalized recommendation refer to individual user profiles, which may be based on persistent or ephemeral data. Consequently they offer recommendations adapted to the individual user.

Figure 3.4 gives an overview of varying degrees of personalization of recommendations regarding (1) the target of recommendations, (2) the typical recommendation method(s) applied, (3) the characteristical data acquirement method, and ( 4) the deployment of user-profiles.

*General recommendations* are suggestions that are given to all users of a recommender system. Typical recommendation methods are statistical sun1IDarization (e.g. Top sellers of all customers of an e-commerce application) and manual selection. Usually no user-specific information is necessary to give this kind of recommendations. As a consequence a user profile is not deployed.

Figure 3.4: Degrees of personalization

Group-specific recommendations are tailored towards a group of users. Usually statistical summarization is applied to generate recommendations for each group. Data acquirement usually takes place by explicit user interrogation (e.g. by offering fields of interest the user can specify, asking for demographic data). If the number of groups the users are segmented into is small, manual selection is also a possible alternative. Personalized recommendations with a short-term perspective are suggestions adapted to the individual user. However, a persistent personalization approach is not pursued. This is suitable, if a authentication of the user is not possible or desired. In e-commerce applications short-term personalized recommendations are often based on items in the virtual shopping basket. Based on this items, complementary items may be recommended to increase cross-sales. Product association rules may be used for this purpose. Ephemeral personalization regularly uses user-profiles to store user-related information. Albeit the profile may be discarded after the user quits the interaction session. Personalized recommendations with a long-term perspective are also adapted to the individual. Information filtering methods in conjunction with persistent user-profiles are typically used to achieve long-term personalization.

Figure 3.5 shows a categorization of recommendation methods based on the personalization criterion. Personalized and non-personalized recommendation methods as well as their advantages and disadvantages are discussed in detail in the following sections.

Figure 3.5: Classification of recommendation methods

#### **3.5.1 Non-Personalized Recommendation Methods**

Non-personalized methods do not adapt recommendations to the user. Hence, all users get identical recommendations. Non-personalized recommendation methods generally require little (statistical summarization) or no (manual selection) computational power. In regard to privacy these methods a.re le.ss problematic, because mapping tastes, preferences, individual characteristics etc. to individual users is not necessary for the recommendation proce.ss.

*Manual selection* refers to the creation of lists of items to recommend by editors, critics, artists and other experts. These lists reflect the personal interests, tastes, preferences and objectives of these specialists and a.re made available to the community. These lists are regularly supplemented by text comments for the individual items to get a better w1derstanding of the recommendations. This method does not require any machine computation at all. Manual selection is a traditional fo1m of providing recommendations and has been used by magazines, newspapers etc. for a long time. By nature, manual recommendations are prone to bias, because they rely on a single persons preferences [SKROl]. However, because they are based on the opinion of experts they may offer deep insights to the items, especially when recommendations are accompanied by high quality text comments. Some e--stores encourage "normal" customers and commwlity members respectively to create manual recommendation lists (e.g. "Listmania Lists" at Amazon.com). Links to specific customer generated lists may be displayed while browsing the product catalogue, if the current article is a part of these lists.

*Statistical summarization* denotes the aggregation of community opinions and community popularity. Typical examples of these summarizations are the munher of community members, who like or pmchase an item or the arithmetic mean of community ratings. A more complex method is to use *association rules* for recommendation purposes. Association rules may be applied on the shopping basket data (i.e. items purchased on a per-transaction basis) of e-- stores [AIS93, SVA97]. A typical example for an association rules would be the finding, that 80 per cent of people, who bought the book "The Last Juror" by John Grisham also bought the "The Da Vinci Code" by Dan Brown. Association rules consist of three elements: (1) the antecedent (in this example"The Last Juror") (2) the consequent ("The Da Vinci Code") and (3) the confidence factor ("80 per cent"), which expresses the strength of the rule.

Table 3.1 shows a simple example of a customer-item matrix for basket data. The columns include different items, the rows contain the customers. A checkmark indicates, that a certain customer has bought the item.


'Ebl31B~d a e .. as et ata: examp. e o a f customer-item matr ix

The customer-item matrix is transformed into an item-item matrix as shown in Table 3.2 by summing up the individual purchase entries. The result of this transformation is always a symmetric matrix (i.e. entries are symmetric with respect to the main diagonal). In this case the figures in the cells show the absolute number of customers who bought a particular item. For example if a customer browses item E, item A (matrix value: 2) would be recommended in the first place followed by item C (value: 1) and D (value: 1).

Product association rules are generally non-personalized (e.g. every customer,


11bl32Bk **a** e : as et d l f ·t 't t . a.ta: examp. e o a 1 em-1 em ma nx

who browses item E will be recommended item A) but can simply be extended to a low level of personalization by using ephemeral navigation patterns ( clickstreams). In this case the values in the corresponding lines may be aggregated. For example if a customer has viewed item A and is currently browsing to item E he or she will be displayed item C (aggregated value: 3) as a recommendation in the first place. Additionally item D (aggregated value: 1) may be recommended. More complex personalized recommendation methods are explained in the following Section.

#### **3.5.2 Personalized Recommendation Methods**

This Section deals with methods for generating personalized recommendations. Personalized recommendations are adapted to the individual users on the basis of knowledge about their preferences and behavior [AT05]. In the following sections the personalization process is illustrated. A general synopsis of information filtering methods is given, characteristics of information filtering methods are described and information filtering is compared to information retrieval. Finally, collaborative filtering, attribute-based filtering, and rulesbased filtering are discussed in detail.

Providing personalized recommendations constitutes an iterative process that is shown in Figure 3.6 and includes the following four stages [AT05]:

1. Define goals and evaluate apprnpriate personalization approaches: Personalization initiatives should be tied to discrete and quantifiable business goals (e.g. increase cross-sales by 10 per cent). Depending on this

Figure 3.6: The personalization process (adapted from [AT05]).

goals and the general condition (e.g. customer base, characteristics of the offered products and services) appropriate personalization approaches have to be evaluated. The pros and cons of the specific approaches (i.e. information filtering methods) are described below.


4. Measure impact of personalization: The last step includes the measurement of the impact of personalization and adequate responses by adjusting the personalization strategy. Measuring personalization impact serves as a feedback for possible improvements of the whole process. This feedback may help to decide whether to collect additional data, build better user profiles, develop better recommendation algorithms or improve the information delivery and presentation [AT05].

#### **3.5.2.1 Synopsis of Information Filtering Methods**

Recommender systems apply information filtering methods to deliver personalized recommendations. Information filtering systems share the following characteristics [BC92]:


filtering process (e.g. jllllk mail filter). In the latter the user sees only the extracted data (e.g. "yom recommendations" at amazon.com).

However, these characteristics are not exclusive to infonnation filtering systems and are also valid for information retrieval systems, which makes it necessary to further distinguish information filtering from information retrieval along the following aspects (BC92, HSSOl]:


• Privacy: Because information filtering systems apply user profiles which may contain sensitive personal data, it is highly concerned with privacy issues, which are mostly of no interest to information retrieval systems.

#### **3.5.2.2 Human Approaches towards Information Filtering**

Based on organizational studies, Malone et al. identified three basic filtering approaches for persons: cognitive, social and economic filtering [MGT+87]. These concepts of human approaches towards information filtering are incorporated into information filtering systems. The characteristics of these approaches are [MGT+87, HSSOl]:


the executive summery instead of the whole report, because his or her workload is high, economic filtering is utilized.

#### **3.5.2.3 Collaborative Filtering**

Collaborative filtering is an approach which applies *similarities between users' tastes and preferences* for recommendation purposes. The basic idea behind collaborative filtering approaches is that the active user will be recommended items, which other users liked in the past (user-to-user con-elation) (SKR0l).

The term collaborative filtering was *first used in literature* by Goldstein et al. (GNOT92). This paper describes "Tapestry", a document filtering system developed at the Xerox Paolo Alto Research Center, which used collaborative filtering to reduce information overload. Tapestry enabled the user to annotate documents (e.g. e-mails, NetNews articles) with text comments and ratings (explicit approach) but also used implicit feedback (e.g. reply to an e-mail as an indicator for relevance) for recommendation purposes. The tapestry system suffered from two problems. Firstly, a small number of users used the system. Because of the absence of a critical mass of users most of the documents were not annotated and hence could not be used for recommendations. Secondly, Tapestry required the user to describe the filtering needs by a complex SQLlike language. This was a hindrance for users to operate the system [ME95). Other early implementations of collaborative filtering systems were Grouplens, Ringo and Video Recommender.

In literature, the distinction between *active and passive collaborative filtering systems* can be found [ME95, Run00). In active systems users actively recommend items to other users (push communication). Active collaborative filtering closely mimics the common practice that people recommend interesting items to other people of their social network (e.g. friends or colleagues). Active collaborative filtering systems supp01t this process by providing information systems as communication tools. Active collaborative filtering requires the user to know interests and preferences of other users. Hence, active systems are of limited scalability. Because of this shortcoming, e-commerce applications regularly apply passive systems for recommendation purposes. In passive systems,

the user does not actively recommend items to other users. A direct communication between the users is not necessary. Passive collaborative filtering uses automated information systems in which people provide recommendations as inputs. These inputs are aggregated and directed to appropriate recipients by the system automatically (SV99). Consequently, passive systems are also referred to as automated collaborative filtering systems.

Table 3.3 illustrates the basic idea of (passive) collaborative filtering based on a simple example. It shows a sample user-item matrix, in which preferences are measured on a binary scale. A "+" indicates that a user liked the item. A "-" means that the user does not like the item. An empty cell indicates, that the user has not rated the item (missing value).


Table 3 3 C : o a 11 b orat1ve filt erm . : examp 1 e o a f user-1 ·t em matrix

Let's assume recommendations are given tho user E. User Eis very similar to user B, because both liked item A and item C and disliked item D. Because user B also liked item E, item E will be recommended to user E in the first place. User A is less close to user E {both liked item A, and disliked item D). Hence, item B could be recommended to customer E additionally. Between user E, user C and user D are no similarities at all. Consequently preferences of user C and user D are not used to give recommendations for user E.

Different statistical methods or machine-learning techniques are applied to calculate the similarity between users. *Memory-based techniques* directly compare users against each others (similar to the example above). They operate over the entire user-item matrix using *statistical methods* to perform similarity measw-es between the users. Correlation-based approaches use the Pearson correlation coefficient ("correlation-based') to determine the similarity between users [RJS+94, SM95, Paz99]. Other memory-based methods use the cosine ("cosine-based") [BHK98, SKKR00) to calculate the proximity between

users. In contrast *model-based approaches* use the users' historical rating data to derive a model. This model is used to make predictions, how the individual users will like certain items. Various *machine-learning techniques* - including Bayesian networks [BHK98], neural networks and latent semantic indexing [FD92] - are used to generate recommendations [Bur02]. However, the latter two techniques typically do not rely on user-ratings solely. Additionally they include attributes of the items (i.e. text documents) in the recommendation process. Hence, they can not be regarded *as* "pure collaborative filtering systems".

The typical *application domain* of recommender systems based on collaborative filtering is to suggest items, whose central characteristics and qualities can not properly measured with "objective" criteria (e.g. books, movies, music) [Run00]. Hence, this items are highly subject to personal taste and preferences.

In order to give reasonable recommendations, correlations of preferences have to exist between users and items. This means that certain groups of users with similar preferences for certain groups of items are given. Collaborative filtering requires a *sufficient number of users* ( "critical mass") and an adequate number of known preferences (i.e. ratings of items) stored in user-profiles to give reasonable recommendations. Because collaborative filtering is based on ratings of a community, it employs human judgement. Thus, it enables the exchange of human knowledge between a large number of people without the requirement of knowing each other personally. This makes collaborative filtering a very powerful approach for recommendations.

In contrast to attribute-based filtering (see Section 3.5.2.4) collaborative filtering systems can give recommendations for items, which have no *"objective" commonalities* in terms of attributes with items the user liked in the past. This may lead to very innovative recommendations from the users' perspective. In fact, the recommendations are founded on relationships between the users of the recommender system, hence similarities between item characteristics are not necessary. For instance, a collaborative filtering systems may recommend a book to the active user because of his past ratings of music or movies. This would be hard to achieve with attribute-based or rules-based systems, because music and books generally have different attributes (an exception would be if a songwriter also works as an book-author; in this case an attribute-based system could recommend books written by the songwriter, because they share the attribute "author").

Collaborative filtering approaches are subject to some *limitations* [BS97, Run00, SKKR00]. The *new user problem* refers to the challenge of giving accurate recommendations to new users. Because the preferences of new users are unknown it is impossible to make appropriate recommendations. An approach to address this problem is to use non-personalized recommendation methods ( e.g. manual selection and association rules) until sufficient preferences are gathered from the user.

The *new item problem* reflects the hindrance to make recommendations for items, which have not been rated by the community. This is usually the case, when new items are added to the database. Because pure collaborative filtering systems solely use community ratings instead of item attributes for the recommendation process, new items can not be recommended [AT03]. Possible solutions are to use non-personalized methods or to combine collaborative filtering with attribute-based filtering ("hybrid approaches") [Bur02, SPUP02]. However the later requires that the object can be reasonably described by objective criteria. These approaches may be accompanied by incentive progranIB to get ratings for new items ( e.g. to offer vouchers for users who write text comments and add ratings to iteillS, which have not been previously rated).

*Rating sparsity* means that the number of given ratings is usually very small compared to the number of iteillS, which may be recommended. This may occur, when the number of users is too small (absence of critical mass of users), when the underlying database of itelllS is rapidly changing or when the users are "too similar" (i.e. all users like and rate the same small set of iteillS). These phenomena lead to a high number of "missing values" [Rm100] in the user-item matrix and consequently reduce the quantity ( "reduced coverage'' because products with no ratings can not be recommended) and quality of recommendations. To address this problem hybrid-approaches may be used. For i11Stance, if the user-profile includes demographic data ( e.g. gender, age, education), this information may be used to find similar users not solely based on similar ratings of iteillS but also on demographic compliance ( "demographic filtering") [Paz99].

The *unusual user* refers to a user, whose tastes are very different from the rest of the population. Hence, it is impossible to find any "nearest neighbors" (i.e. like-minded users) to derive recommendations from their ratings. Hence, the quality of recommendations for this kind of user are poor.

Collaborative filtering systems pe1form complex mathematical operations over large amounts of data. For the user it is hard to understand, why a certain item is recommended. This is called the *"Black-Box problem".* A possibility to enhance transparency of the recommendation process is to display explanations (see Section 3.2).

*Scalability problems* may arise when collaborative filtering methods are used, because with this technique computation grows with the number of users and the number of items. In e-commerce applications these systems are challenged with millions of users and items. Consequently serious scalability problems may occur [SKKR00]. This is especially the case when memory-based algorithms are used. As mentioned above, memory-based algorithms operate over the entire database (which contains the user-item matrix) to give recommendations [BHK98]. Hence, this algorithms are prone to scalability problems. In contrast model-based approaches use the database to estimate parameters of a model in advance. This model is used to give recommendations to individual users after the calculation of the model parameters. Thus, it is not necessary to access the whole database while giving recommendations to the user. Consequently, model-based approaches outperform memory-based algorithms but may show a lack of accuracy, especially when the database is frequently changing [BHK98].

Collaborative filtering systems *disregard product attributes* for recommendation purposes, even when they are of high relevance. Pure collaborative filtering systems are not reasonably applicable, when the "objective" criteria of the recommended items are dominating the user's preferences. For instance in the application domain of personal computers, objective attributes (e.g. performance data) have a strong influence on the buyers decision malting process. The impact of subjective criteria (like the user's brand affinity) on the buying decision may still be given, but is usually of less importance. Consequently, the quality of recommendations based on collaborative filtering techniques may be considered as poor, because the user's requirements regarding these objective attributes are not taken into consideration. In addition, associations of items based on similarities between item characteristics can not be discovered by collaborative filtering systems. For example, a user likes fihns directed by Robert Rodriguez. A collaborative approach can not recommend all movies, music or books by Robert Rodriguez, because the attributes ( e.g. "directed by", "composed by" and "written by") and the corresponding relationships are not modelled in pure collaborative filtering systems.

#### **3.5.2.4 Attribute-Based Filtering**

*Attribute-based filtering* is an filtering technique, which uses similarities between items for recommendations. This fundan1ental assmnption is, that a user will like items similar to the ones he or she liked in the past [BS97].

In attribute--based filtering systems, the interest of a user is determined by the associated featmes of items. Hence the term *''feature-based''* approaches is also used for such systems [Run00]. Because the basic idea of this method is an outgrowth of information filtering research and was initially applied on textual documents, the term *"content-based filtering"* is a further term found in literature to describe such systems [BS97, Bur02, HSSOl]. Strictly speaking, content-based approaches are a subclass of attribute--based filtering systems, where the application domain is textual documents. These documents are described by a restricted number of attributes of the content ( e.g. characteristic words) [SPK00].

Similar to collaborative filtering, attribute--based filtering approaches employ a long-term user-model to learn and store user-preferences. In contrast to collaborative filtering, the interests of the user are not determined by comparing the similarity of the user to other users. Instead the interests of the user are derived from the attributes of the items, the user has already rated. Hence, attribute--based filtering systems generate recommendations based on a user-profile built up by analyzing the attributes of items which the user has rated in the past [BS97].

When designing a content-based filtering system two problems have to be addressed [Paz99]:


Table 3.4 illustrates a representation scheme as described above in conjunction with a user profile. In this simple and fictional example books on e-commerce are represented by four keywords. A checkmark indicates that the term corresponding term occurs in the description of the book. A "+" in the column of "User A" means, that the user was interested in the book. A"-" indicates, that the user was not interested in the book. Because Book E and F are unknown to the user, they can be used for recommendation purposes. For example Book E might not be of interest, because in the past the user was not interested in books, which dealt with E-Branding. However, he or she might be interested in "Book F", because it covers topics the user is interested in.

Applying attribute-based filtering requires two preconditions: ( 1) The items can be described by "objective" criteria and (2) there must be a significant


T **a** bl e 3 4 A : ttn "b ute-b *as* ed fil term1r

coherence between these criteria. and the global preferences of the users of attribute-based recommender systems [Rw100].

Consequently, attribute-based filtering systems are well suited for domains where subjective tastes are not dominating the selection process and judgments a.re merely based on "hard-facts" (e.g. technical products). A typical example are digital cameras, which can be described with technical data. However, subjective criteria. (e.g. design, brand-attitude} might still play a considerable role in the purchase decision process. Attribute-based filtering systems have limitations in the incorporation of these subjective criteria.

In contrast to collaborative filtering, attribute-based filtering methods do not depend on ratings of other users than the active user. Hence, attribute-based filtering systems are faster applicable than collaborative filtering systems, because building a "critical mass" of users is not crucial for the deployment of attribute-based systems (Run00].

A fmther advantage of is the structured representation of the attributes of items. Consequently, these meta-data could be used for purposes that go beyond attribute-based filtering. For example, the search for specific attributes is easy to implement (e.g. "show all books written by Umberto Eco"). Rulesbased filtering approaches (see Section 3.5.2.5) can further be applied, when stmctured meta-data of items are already existent.

However, attribute-based systems are prone to some limitations. *Limited content analysis* refers to the fa.ct, that attribute-based systems are limited by the attributes that are explicitly linked to the items these systems recommend [AT03]. Depending on the domain, these features can be extracted automatically or have to be assigned by hand. As mentioned above, information retrieval offers a variety of methods to extract features of textual documents automatically. However, in other domains (e.g. multimedia-data) automatic feature extraction is much more complicated [BS97]. The assignment of attributes by hand is a time consuming task, which is often not practical due to limitations of resources [SM95]. Depending on the application domain a fmther problem with limited content analysis may be that two items with the same associated attributes may be indistinguishable. This may be of no concern when the two items are equivalent ( e.g. technical products with the same specifications). However, if attribute-based filtering systems are applied on textual documents ( "content-based filtering") a problem might occur. Textual documents are usually represented by the "most important keywords". Consequently, a well-written article can not be distinguished from a bad one, if the same terms are used [SM95].

*Over-specialization* is a further shortcoming of attribute-based systems. Because this kind of filtering system can only recommend items that score highly against the active user's profile, the user is limited to get recommendations of items that are similar to those already rated [BS97]. Consequently, the recommendations of attribute-based systems may not appear as "innovative" to the user compared to recommendations based on collaborative-filtering algorithms. In some fields of application, items that are to similar should not be recommended ( e.g. articles in different newspapers, which describe the same event). Hence in some cases it may be sound to filter out items which are too similar to the ones the user has rated or seen before additionally [AT03].

Similar to collaborative filtering, attribute-based filtering systems also face the *new user problem.* If the number of ratings in the user-profile is insufficient, the system is not able to give accurate and reliable recommendations.

#### **3.5.2.5 Rules-Based Filtering**

Rules-based filtering is an approach that employs *business-rules* for recommendations. In this context, rules describe on-line behavioral activities of the

users [ATOl). In general, rules-based approaches can be designed stereotypical or personalized [KSS03). In stereotype rule-based filtering approaches, the individual user is assigned to a group of similar users. For filtering pw-poses, the identical set of rules is used on each member of the group. In contrast personalized ntles-based filtering systems apply an individual set of ntles for each user [KSS03). Consequently, the degree of personalization is higher with the latter approach.

Rules-based approaches are widely used in the field of expert systems for knowledge representation purposes [Jac98). Generally rules may be described in the following form: IF {predicate} THEN {result}. In personalized rulebased filtering approaches a user profile contains a *set of rules,* that expresses the preferences of an individual user [KSS03). For instance: IF {book.abstract contains "Macroeconomics" and book\_year\_oLpublication not less than 1995} then { user ..relevancy = "very high"}.

The main task of rule-based approaches is to *discover suitable rules.* In general, finding appropriate rules is accomplished with human experts ( e.g. a marketing manager). Consequently, the effort for employing rules-based approaches tends to be higher compared to collaborative and attribute-based filtering methods due to the involvement of *human expertise.* 

Figure 3. 7 shows a *structured approach* towards the rule discovery process. In order to get "truly" personalized recommendations, rule discovery methods are applied to the data of every single user. The process of discovering rules could be divided in two phases: (1) data mining and (2) validation of the rules [ATOl).

In the first step, *data mining methods* are applied on the user data to generate a large set of rules. Many of these rules are trivial, spm-ious and not relevant in the given application domain [ATOl]. Hence step two, i.e. *role validation,* is an important issue with this approach to get high-quality recommendations. Because of the sheer number of rules and users in e-commerce applications it is impossible to validate each rule for an individual customer by a domain expert. Consequently, rule validation is not pedormed separately for each user, but for all users at once by applying rule validation operators. Because there are many similar or identical rules across different users, validation effort can

Figure 3.7: The rule discovery process [AT0l]

be significantly reduced. In the end, the accepted rules form the profile of the individual users. For a detailed description of this process see [AT0l].

One of the major drawbacks of rules-based filtering is the relative *static* nature of this approach. In contrast to collaborative filtering, changes in the taste of the user-population is reflected over the time due to the permanent rating of items by the users. However, in rules-based approaches the rules stay the same until a new discover and validation process is initiated. Because this process uses human expert knowledge, the effort of updating the rules is much higher compared to the "automatic" collaborative-filtering approach.

### **Chapter 4**

## **Research Model, Hypotheses, and Methodology**

This chapter deals with the research model. In the first step the problem statement of the work is defined. Based on that, the research questions and the research model are elaborated. Thereafter, the hypotheses are summarized. The chapter ends with a section that deals with methodological aspects.

#### **4.1 Problem Statement**

The majority of research literature regarding recommender systems deals with this topic from the viewpoint of computer science. The focus is on the underlying algorithms for generating recommendations [KSS03, SKKR00, 8S97, Bur02, SVA97, Run00]. The existing research concerning the marketing perspective ( e.g. the influence of recommendations on consumers decisions) is still scarce [SN04, HK04, HM03, CLA+o3, HT0O].

The book strives to identify the underlying psychograph.ic factors of conswners that determine: (1) the interest in *personalized recommendations,* (2) the interest in engaging actively in virtual communities of transaction located at online purchase environments by *submitting product-related ratings and comments,* 

and (3) the interest in *product-related opinions of other consumers* in virtual communities.

Virtual communities a.re important for recommendation applications, especially if collaborative filtering is applied for recommending products and services. As mentioned in Section 3.5.2.3, collaborative filtering is an approach which applies *similarities between 'USers' tastes and preferences* for recommendation pmposes. The basic idea behind collaborative filtering approaches is that items a.re recommended to the active user, which other users liked in the pa.st (user-to-user correlation) [SKROI]. Especially when using explicit data acquirement, it is important to have a lively community organized at the online pmchase environment in order to learn preferences of consumers for recommendation purposes.

However, when collaborative filtering is applied, the following problems arise (for a detailed description see Section 3.5.2.3):


Consequently it is of interest, which psychographic factors of conswners are tangent to the problem areas mentioned above. The following section deals with the research questions based on these problem areas.

### **4.2 Research Questions and Model**

The central research question of the book is: *Which psychographic factors are of major importance for the acceptance of online product recommendations*  *and the commitment to participate in the virtual community of an e-vendor by submitting rotings and comments of products?* 

The author tries to address this question by applying the *opinion leadership theory* in the context of online book recommendations. The author has chosen books as the product class because of the following reasons:


*Opinion leadership* is a well-established and well-researched concept in marketing [BME0l, MG95]. The term "opinion leadership" was introduced to scientific debate by Lazarsfeld et al. in 1944 [LBG44]. The study of the 1940 presidential election examined the influence of relatives, friends, and coworkers on voting decisions. The concept was applied to the field of consumer decisions by Katz and Latzarsfeld in 1955 [KL55]. Empirical evidence of the importance of opinion leadership was fostered by King and Swnmers in 1970 [KS70].

In the field of consumer decisions opinion leadership is understood as the *exertion of an unequal amount of influence* by consumers in the purchase behavior of others [FGE96]. In general, opinion leadership stimulates interpersonal communication ("word-of-mouth"). One aspect of this process is that opinion leaders tend to give recommendations to other consumers ( "advice giving word-of-mouth"). With the application of recommender systems e-commerce vendors try to mimic or support this process by the use of information systems. Hence, it seems suitable to apply the opinion leadership concept to get a better W1derstanding of online recommendations and commWlity activity in e-commerce environments.

Consequently, the question arises which underlying factors determine opinion leadership. Marketing literature has identified *involvement with the product category as* an important factor of opinion leadership [RRS98, FP87, RD71]. Product involvement is often viewed *as* the long-term interest in a product class based on the centrality to important values, needs, or the selfconcept [Blo81].

Figure 4.1 summarizes the already empirical tested *background theory* of the book. Product involvement positively affects opinion leadership and opinion leadership itself has a positive influence on word-of-mouth [RRS98].

Figure 4.1: BackgroWld theory of the book

The research model shown in Figure 4.2 adapts the basic research model *towards e-commerce applications* and includes the *interest in online-product recommendations.* Word-of-mouth is specified as the interest to contribute product-related comments and ratings to the virtual commWlity of an evendor. Acceptance of recommendations in general is defined by the interest in receiving personalized online recommendations by an e-commerce application. Further, the opinion seeking concept is added to the model. Opinion seeking occurs, when individuals search out for advice from other consumers when making a purchase decision with respect to a certain product class [FGE96].

Because some inadequacies regarding the involvement variable have been identified in literature [RRS98], the Wlidimensional approach to product involvement is substituted by the *multifaceted construct of product involvement* proposed by Kapferer and Lament. According to these authors, involvement is a multifaceted construct along five dimensions [KL86]. It consists of the perceived importance and risk of the product class, the subjective probability of

Figure 4.2: Research model

making a mispurchase, the symbolic or sign value, the hedonic value of the product class, the hedonic value of the product class and the interest in the product class.

The author assumes that the symbolic or sign value and the hedonic value facets of product involvement influence the opinion leadership behavior. F\trthermore, a positive relationship between the risk of a mispurchase facet and opinion seeking is asswned. Additionally, it is hypothesized that opinion seeking behavior has a positive effect towards the interest in reading productrelated comments and ratings. Finally, it is assumed that the participation in a virtual community (i.e. reading and submitting product-related comments and ratings) has a positive influence towards the interest in personalized online recommendations.

Figw-e 4.3 presents the extended research model, where the influence of further psychographic and sociodemographic factors is examined. Domain-specific innovativeness reflects the tendency to learn about and adopt new products ( or innovations) within a specific domain (i.e. product class) [GH91]. It is asswned that *domain-specific innovativeness* has a positive influence on the interest in recommendations. *Impulse buying tendency* is a further psychographic determinant found in the extended model. It refers to the degree to which an individual is likely to make unintended, immediate, and unreflective purchases

Figure 4.3: Extended research model

(i.e. impulse purchases) [WJB97]. A positive influence towards the interest in personalized online recommendations is expected. A further psychographic factor found in the extended model is *skepticism towards advertising.* This factor is defined as a general tendency toward disbelief of advertising claims [0S98]. Because personalized online recommendations can be understood as personalized kind of advertising, it is asswned that skeptic persons have a lower interest in recommendations. The influence of *privacy concerns* and *experience with online shopping* is also investigated in this extended model.

In addition, the influence of demographic factors (e.g. age, gender, income) are investigated. Table 4.1 swnmarizes the *hypotheses that are derived from the extended research model* and are investigated in this book.

Besides the influence of psychographic and sociodemographic determinants on the acceptance of recommendations and community activity, the following research questions a.re addressed in this book by means of exploratory research:


Table 4.1: Research hypotheses



*sources (e.g. other consumers, critics, trusted third parties) for the decision process of the consumers?* 


### **4.3 Methodology and Research Design**

In the book a *quantitative research approach* is applied. As mentioned above, the research model is tested in the context of book recommendations. Hereby, consumers were asked to answer a standardized web-based questionnaire. The research model shown in Figure 4.2 (see Section 4.2) is verified by the application of *structural equation modeling.* Further, the psychographic determinants shown in Figure 4.3 in the lower box are tested by a *regression analysis,* because including all this factors in a structural equation model would have been overly complex. The demographic factors are verified by a regression analysis and Mann-Whitney tests (for gender-specific differences).

The following multi-item, self-report scales are used for the measurement of the psychographic factors:


The original scales were translated into German. To avoid adulteration, the measures were translated by the author, retranslated by an independent translator and finally verified by an independent native speaker. These factors are measured along a seven-point Likert scale ranging from "totally disagree" to "totally agree".

The rest of the scales used in the research model (i.e. interest in personalized book recommendations, interest in writing book-related reviews, interest in reading book-related reviews, experience with online shopping, and privacy concerns) were developed by the author of the book. In accordance with the scales taken from literature, a seven-point Likert scale was used for measurement.

In Figure 4.4 the research design of this book is illustrated. In the first step a literature research was conducted and the problem statement defined. Based on that, the research model and the corresponding hypotheses set forth in this chapter were elaborated. The next stage included the development of the web- based questionnaire. As mentioned a translation- and retranslation-process was initiated to reduce adulteration due to language aspects.

After a pre-test phase (which included six persons) the first survey was conducted. This sw-vey was performed in collaboration with the Austrian bookseller A&M Andreas & Dr. Millier Verlagsbuchhandel (www.aurn.at). According to the Austrian Internet Radar A&M is on position nwnber eighteen of Austrian web-sites with respect to the range of coverage. 16% of the Austrian Internet users in the sample have visited this web-site "within the last four weeks" (starting from the time of questioning, survey period: 2005-09-15 to 2005-12-15, n=5000) [AIR05]. For comparison, the world's biggest online

Figure 4.4: Research design

bookseller Amazon is on position number seven of Austria's most visited websites with a coverage of 34% (AIR05]. The survey was conducted from July 8th 2005 to September 2nd 2005.

In the next step, an exploratory factor analysis was conducted. The goal of this analysis was to shorten the questionnaire for the next survey and to determine the items that should be included in the structUl'al equation model. To avoid fitting the model to the data ( which would happen if the structural equation model was calculated on the whole dataset), the dataset was split. 20% of the data was used for an exploratory factor analysis. In this context, the three items of the scale with the highest factor loadings were chosen to be included into the structmal equation model ( the calculation was based on the remaining

80% of the data) and the construction of the questionnaire for the follow-up survey (i.e. the other items with lower loadings were removed from the original questionnaire).

The follow-up survey was conducted in cooperation with the biggest Austrian Internet service provider Telekom Austria AG (www.aon.at). According to the Austrian Internet Radar the Telekom Austria AG is on position number three of the Austrian web-sites with a coverage of 45%. The reasoning behind the follow-up survey was to analyze the two samples in regard to the differences. The results that stem from survey posted at the web-site of a bookseller are clearly of highest relevance for the purposes of this book, especially with respect to the composition of the sample. However, a survey posted at the web-site of an Internet service provider should be a good supplement, because the resulting sample is thought to represent the Austrian Internet population as a whole. To sum things up, the author assumes that the results derived from the first survey stand for typical Austrian "online shoppers with an interest in books", whereas the results from the second survey stand for the "general Austrian Internet population" .

### **Chapter 5**

### **Results**

In this chapter, the empirical results are presented. In Section 5.1 the descriptive results (i.e. results that are not related to the hypotheses and the research model respectively) are illustrated. In this context, the sample size and demographic data are described. Further, results in regard to Internet usage, online shopping, product recommendations, ratings and comments are depicted. Section 5.2 deals with the the verification of the research model and the hypotheses. A factor analysis, a structural equation model that tests the psychographic hypotheses, and regression models that verify further psychographic and sociodemographic hypotheses are depicted.

As mentioned in Section 4.3 two independent surveys were conducted. The first survey was made in cooperation with the Austrian bookseller A&M Andreas & Dr. Mi.iller Verlagsbuchhandel (www.aum.at). Results of this survey are hereinafter referred to as AUM. The second survey was conducted in cooperation with the Internet service provider Telekom Austria AG (www.a.on.at). For results of this survey the acronym AON is used in the subsequent sections.

#### **5.1 Descriptive Results**

In the following, the descriptive results are set out. Firstly, a presentation of the san1ple size and the demographic data of the two surveys is undertaken.

Hereinafter, results of the Internet usage followed by descriptive results in respect to online shopping, online product recommendations, as well as ratings and comments are shown.

#### **5.1.1 Sample Size and Demographic Data**

The smvey AUM was conducted from July 8th 2005 to September 2nd 2005. In total 682 participants filled out the questionnaire on the booksellers website. Smvey AON was conducted from November 21st 2005 to December 5th 2005. In this survey 396 respondents were involved.

Figure 5.1: Sample description: Gender


Figure 5.1 and Table 5.1 compare the two surveys in respect to the gender of the survey participants. 34.5% of the study participants of survey AUM (i.e. the survey conducted in cooperation with the bookseller) are male and 65,05% are female. Table 5.1 points out that 104 respondents refused to specify their gender. In this context the general proportion of male and female in the Austrian Internet user population is of interest. According to the GfK Online Monitor for the 3rd quarter 2005, 55% of the Austrian Internet user are male and 45% are female [GfK05]. As shown in Table 5.1, women are overrepresented by 20.05 percentage points in the sample AUM. In contrast in the survey AON (i.e. the survey posted on the web-site of the Internet service provider) woman a.re underrepresented by 11.12 percentage points. In this survey 66.12% of participants are male and 33.88% are female.

Figure 5.2: Sample description: Age pattern



Figure 5.2 shows the age pattern of the two samples. In the survey AUM the youngest participant is 13 years old, the oldest has an age of 78. The arithmetic mean of age accounts for 36.2 yea.rs. More specifically, the arithmetic mean for men is 40.0 years, the arithmetic mean for women is 33.7. In the survey AON the minimum age is 12 years and the maximum is 69 years. The arithmetic mean in this survey accounts for 39.85 years. In this sample, the arithmetic mean for man is 42.5 years and the arithmetic mean for woman is 34.4 years. As shown in Table 5.2 the differences in respect to the age {RESPOAGE) of participants between the two samples a.re significant (o: = 0.05). The nonparametric Mann-Whitney test was performed, because normal distribution of the variable age was not given in the two samples.


11bl 53 : S a e a.mp 1 d e escnpt10n: A ,ge pattern nv~ nv .. , •-•u **Au•trian** 

Table 5.3 compares the age patterns of both surveys to the age pattern of the general Austrian Internet population in the 3rd quarter of 2005. As shown in both surveys young and elderly people a.re undenepresented. Additionally the deviation from the general Internet population in Austria is depicted in percentage points for ea.ch class [GfK05].

The boxplots in Figure 5.3 swmnarizes the distribution of age with respect to the gender and highlights that women in both samples a.re on the average younger than ma.le participants.

Figure 5.4 illustrates the occupation of the respondents of both surveys. In survey AUM the majority (46.8%) are white-collar employees, followed by pub-

Figure 5.3: Sample description: Age by gender

lie servants (13%) and blue-collar employees (7.9%). 111 survey participants rejected to answer the question regarding the occupation. In survey AON white-collar employees are also the largest group (39,7%). To be consistent with study AUM, public servants constitute the second largest group (12.5%). In contrast to the first survey, retirees are the 3rd largest group (9.4%). Further, the percentage of blue-collar workers is equal to the self-employed people (8.1%). In this survey 99 persons did not answer the question regarding the occupation.

The educational levels of the respondents are depicted in Figure 5.5. In survey AUM, the two largest groups are survey participants with a final apprentice examination (30.1%) as well as respondents, who have attended a secondary school and received a diploma qualifying for university entrance (29.4%). 14.3% of the respondents possess an university degree. People who attended primary school solely account for 10.6%. 15.5% attended other educational institutions. 108 respondents did not answer the question regarding the educational level. In survey AON the overall educational level is slightly higher. The leading group consists of people, who attended a secondary school and received a diploma qualifying for university entrance (34.1%) followed by people with a final apprentice examination (30.1%). In contrast to the first

Figure 5.4: Sample description: Occupation

study people with a university degree form the third largest group. In survey AON 94 persons left out the question regarding their educational level.

Figure 5.6 compares the number of people in household for both surveys. 26.4 % of the people in survey AUM live in households with 3 persons. 25.3% live in households with 2 persons followed by 21 %, where the number of people in household is 4. Single households account for 12. 7% in this survey. The largest household in respect to the number of persons is 8. In survey AON households with two persons (27.2%) are in the majority followed by households with 4 persons (25.5%). The leading group of survey AUM (i.e. households with 3 persons) are the third largest group in survey AON (21.1%). Interestingly, single households are less frequent than households with 5 persons. They account for 10.5%, whereas households with 5 persons account for 12.2%. In accordance with the first survey, the maximum number of people in a. household is 8. In survey AUM 148 persons did not specify the number of persons in household. In survey AON 102 values are missing.

Figure 5.6: Sample description: Number of people in household

The monthly household income of the respondents is shown in Figure 5.7. In general the respondents where very reluctant to give this information. In smvey AUM 47.80% of the overall sample (i.e. 61.16% of the valid responses)

Figure 5.7: Sample description: Monthly household income

explicitly refused to specify this by marking the relevant field (i.e. "not specified") in the questionnaire. In addition, in 149 cases the respondents did not fill out this question at all. The largest group (14.37%) that specified the monthly household income has between 1001 € and 2000€ at disposal. The monthly household income of the second largest group (6,74%) is between 2001 € and 3000€. 6,3% of the respondents have less or equal than 1000€ per month. The situation in survey AON is nearly identical. The majority (47%) of the respondents marked "not specified" in the questionnaire. In accordance with survey AUM people with a monthly household income between 1001 € and 2000€ form the largest group that has specified the income (21.4%). In contrast the third largest group are people with an income between 3001 € and 4000€, whereas in survey AUM the third group that specified the income are people with an income below or equal 1000€. In the second survey 111 values regarding the monthly household income are missing.

#### **5 .1. 2 Internet Usage**

The following section deals with the presentation of descriptive results in respect of the Internet usage.

Figure 5.8: Sample description: Internet usage in years

As illustrated in Figure 5.8 in both surveys the majority has used the Internet for more than 4 years. In survey AUM this group accounts for 80.2%. 24.1 % of the users in this survey have browsed the Internet for between 2 and 4 years, followed by 7.4%, who have used the Internet for between 1 and 2 years. 2.3% are relatively new to the Internet. They have experienced the Internet for less than 1 year. 102 respondents did not fill out their experience with the Internet in terms of years using it. In survey AON 80.2% have used the Internet for more than 4 years. 15.2% used the Internet for between 2 or 4 years. The both last groups in this survey account for 2.3% each. 93 values regarding Internet experience in years were missing in survey AON.

The average time in hours spent surfing on the Internet is illustrated for both surveys in the histograms of Figure 5.9. The boxplots below show the distribution of values in both surveys. In survey AUM, the respondents spent 12.96 hours on average per week on the Internet, whereas in the survey AON the arithmetic mean accounts for 15.33 hours.

Figure 5.9: Sample description: Time spent on the Internet (weekly, in hours)

According to the Mann-Whitney test shown in Table 5.4 the differences **in**  regard to the time spent online (INTEHOUR) between the two surveys are significant *(a* = 0.05). Additionally, gender-related differences of time spent browsing the Internet for both surveys were investigated. As shown in Table 5.5 no significant differences *(a* = 0.05) between males and females in respect of the time spent online were found in the two surveys.

Figure 5.10 illustrates from which places people have access to the Internet. Multiple answers are possible in this question. In sm·vey AUM, 78. 7% of the respondents have access to the Internet from their home. 38.4% may



0.01

Z

Asymp. Sig. (2-tailed)

Table 5.4: Mann-Whitney test: Time spent on the Internet

Table 5.5: Mann-Whitney test: Time spent on the Internet by Gender

Figure 5.10: Sample description: Access to the Internet

utilize the Internet from their workplace. 6.0% have access from educational institutions and 6.2% access the Internet from other places. The situation in survey AON is very similar. 74.5% use the Internet at home, 40.7% use it from the workplace, 10.6% from home, 10.6% access the Internet from educational institutions. 5.8% utilize the Internet from other places.

#### **5.1.3 Online Shopping**

In this section Internet shopping related questions of the surveys are discussed.

Figure 5.11: Sample description: Internet shopping frequency

In Figure 5.11, the Internet shopping frequencies of the two surveys a.re compared. In survey AUM, 25.6% of the survey participants buy online several times per month. The majority (58.0%) of the valid responses purchase online several times per year. 7.1% shop circa one time a year and 9.2% buy less frequent. In 108 cases, the specification of the shopping frequency is missing. In survey AON the situation is nearly identical. 24.0% of the respondents buy several times per month. 42.7% shop several times per year, followed by 7.3%, who acquire products and services online circa one time a year. 12.3% buy less frequent. 96 respondents did not answer this question.

In this context, differences in the shopping frequency in respect to gender,

time spent on the Internet and age were investigated. According to the nonparametric Mann-Whitney test shown in Table 5.6 no significant differences in the shopping frequency (FREQSHOP) were found between men and women in both samples.

Table 5.6: Mann-Whitne test: Internet sho uenc b gender


Regarding the age of respondents and time spent on the Internet a bi-variate correlation analysis (Spearman's rho) was performed. Table 5.7 shows that no significant relationship between shopping frequency and age (RESPOAGE) of the respondents was found. The relationship between time spent on the Internet (INTEHOUR) and buying frequency is of high significance *(a=* 0.05) in both surveys. The correlation coefficient is negative because of the reversed coding of shopping frequency (i.e. 1 means a high frequency, 4 means a low frequency). Hence, people that spent a lot of time on the Internet also have a higher buying frequency, which is pretty obvious.

Table 5.7: Bi-variate correlation analysis: Internet shopping frequency, age and time spent on the Internet


The online shopping frequency of books of the two surveys is illustrated in Figure 5.12. The respondents were asked to answer this question on a seven point Likert scale ranging from never (value = 1) to very frequent (value =

Figure 5.12: Sample description: Internet shopping frequency of books

7). As the bar cha.rt shows, the situation is quite different in both surveys. Although in both smveys the majority buys books '"sometimes" (i.e. 26.3% in survey AUM, 25.0% in swvey AON), in the other categories the two surveys differ substantially. In survey AUM "heavy buyers" (i.e. consumers that buy books more often than "sometimes" a.re dominant, whereas swvey AON exhibits buyers predominantly, who buy books less frequently than "sometimes". These results a.re quite obvious, because survey AUM was posted on the web site of a bookseller, whereas survey AON was posted on the web site of an Internet service provider. In survey AUM the arithmetic mean is 4.48, in survey AON the arithmetic mean of the book-related buying frequency is 3.57. In survey AUM, the number of missing values accounts for 12, whereas in survey AON 8 respondents did not fill out this question.

In Figure 5.13 the shopping frequency of music in the two smveys is compared. In general music is more seldom pmcha.sed than books in both smveys. The arithmetic mean in survey AUM is 3.53 (compared to 4.58 in the book category). In contra.st to the category books, the number of people that buy music less frequently than "sometimes" outweigh the "heavy buyers" (i.e. consumer that buy music more frequently than "sometimes"). Although, the majority

FigW"e 5.13: Sample description: Internet shopping frequency of music

(26.5%) "sometimes" buys music. In contra.st, in SW'vey AON the most people (30.4%) "never" buy music, followed by people who "sometimes" buy music online (26.7%). The arithmetic mean accounts for 3.09 (compared to 3.57 in the book category). The number of missing values in this category is 22 (AUM) and 14 (AON).

As Figure 5.14 illustrates, movies is the least sold product category in both surveys. In survey AUM, the percentage of people that "never" (22.3%) buy movies is nearly equal to the percentage of people that sometimes ( 22. 7%) buy movies. The arithmetic mean in this survey accounts for 3.22. In survey AON, consumers that never buy movies are by far the dominant group (41.6%). The arithmetic mean adds up to 2.58. The number of missing values is 27 in survey AUM and 19 in survey AON respectively.

The shopping frequencies of this three product categories reflect the results of the Austrian Internet Monitor in the third quarter of 2005 [AIM05]. In this study the top 10 of products sold over the Internet a.re presented. The nun1ber one product category is represented by books (37% of the Austrian Internet user have bought a book in the last three months) followed by clothing and

Figure 5.14: Sample description: Internet shopping frequency of movies

shoes. Music is the 5th most sold product (14%), whereas movies a.re on place 8 of the list (9%).

Two Mann-Whitney tests were performed to investigate gender specific differences of buying frequencies. The results a.re presented in Table 5.8 and Table 5.9. FREQBOOK refers to the buying frequency of books, FREQMUSI to the frequency of music and FREQMOVI to the frequency of movies. Interestingly for the product categories music and movies no significant differences between male and female respondents were found in both surveys *(a=* 0.05). On the other hand books a.re significantly more often bought by women in both surveys *(a=* 0.05).

In Table 5.10 a bivariate correlation analysis (Spea.rman's rho) for each product category is shown. In the following interpretation of the correlation analysis a significance level of 0.05 is assumed (i.e.a = 0.05).

In survey AUM, the product category book shows no significant relationship between age and buying frequency of books. Further, no significant correlation between time spent on the Internet and buying frequency exists according to the survey. In contrast, in smvey AON a significant correlation between time

Table 5.8: Mann-Whitney test survey AUM: Internet shopping frequency of books, music and movies by gender


Table 5.9: Mann-Whitney test survey AON: Internet shopping frequency of books, music and movies by gender


spent on the Internet and buying frequency of the product category book is found.

The buying frequency of music has a significant positive relationship with the time spent on the Internet in both surveys. In respect to the age of the respondents no significant correlation was detected in both surveys.

Interestingly, the buying frequency of movies has a significant negative relationship with the age of the respondents. Hence, it can be assumed that as a tendency movies are bought by younger people. Additionally, a significant positive relationship between time spent on the Internet and the buying frequency of movies is found in both surveys.

For the correlation analysis of education and shopping frequencies of the three


Table 5.10: Bi-variate correlation a.na.lysis: Internet shopping frequency of books, music and movies

products types cases with the category "other" were excluded from the dataset to obtain a distinct ordinal measurement scale. As the table shows, in both surveys books are more often bought by people with a higher educational level. For the shopping frequency of music no significant relationship is detected in respect to the educational level of the respondents in both surveys. Regarding movies the results of the two surveys differ. Survey AUM shows a significant negative relationship, whereas survey AON shows now significant differences.

#### **5.1.4 Online Product Recommendations**

In this section descriptive results with respect to online product recommendations are presented.

Figure 5.15 contrasts the percentage of people that have got recommendations in e-commerce applications to people that have never received such recommendations. In survey AUM, 78.6% of the survey participants have already

Figure 5.15: Sample description: Recommendation received

received a product recommendation in an online shop. 21.4% stated that they have never got a recommendation. At the time of the survey the online shop of the bookseller did not employ any kind of recommender system. In the second survey (AON) (i.e. the questionnaire posted at the web site of the Internet service provider), the percentage of people that have been exposed to online product recommendations is even higher (85.4%). Thence, 14.6% of the respondents have never got a recommendation in this survey. The number of people who did not answer this questions adds up to 3 in survey AUM and 12 in survey AON respectively.

In the following buying frequencies of books, music, and movies that were bought because of an online product recommendation are investigated. In the questiom1aire the three questions were designed as filter questions. Hence, these questions were only displayed to people, who answered the question if they have already recieved product recommendations with "yes" (See Figure 5.15).

In Figure 5.16 the buying frequency of books that were bought because of a recommendation is illustrated. In both surveys the group of respondents that "sometimes" buys books because of a recommendation is leading (AUM 27.4%,

Figure 5.16: Sample description: Bought books because of recommendation

AON 24.2%). As shown in the bar chart in survey AUM the respondents that buy books more often than "sometimes" are predominant over respondents that buy less frequently than "sometimes". In sw-vey AON, the situation is contrary. Therein, people that buy books less frequently than "sometimes" outweigh people that buy books more frequently than sometimes because of a given recommendation. The arithmetic mean for this question accounts for 4.00 in survey AUM and 3.23 in sw-vey AON respectively.

Figure 5.17 illustrates the buying frequency of music due to recommendations in online shops. In survey AUM 26.0% of the respondents have "sometimes" bought music because of a recommendation, followed by 24.2%, who have "never" bought music due to a suggestion in an online store. In contrast, in survey AUM the majority {36.0%} has "never" bought music because of a recommendation. 23.3% have sometimes purchased a book because of this. In both surveys, the percentage of people that have bought music less frequently than sometimes outweighs the percentage of people that have bought music more frequently than sometimes. The arithmetic mean in survey AUM is 3.18. In the other survey the arithmetic mean accounts for 2. 75.

Figure 5.17: Sample description: Bought music because of recommendation

Figure 5.18: Sample description: Bought movies because of recommendation

As Figure 5.18 clearly shows that movies are the least frequent product category bought due to recommendations. In both surveys, the majority has "never" bought a movie because of a recommendation (32.2% in survey AUM and 49.6% in survey AON), followed by respondents that buy sometimes books (20.3% in survey AUM and 15.7% in survey AON). The arithmetic mean in survey AUM is 2.94 and 2.32 in survey AON respectively.

Table 5.11: Mann-Whitney test survey AON: Internet shopping frequency of books, music and movies bought because of recommendations by gender


Table 5.12: Mann-Whitney test survey AUM: Internet shopping frequency of books, music and movies bought because of recommendations by gender


As Table 5.11 and Table 5.12 show, the two studies deliver a rather inconsistent picture of differences between male and female respondents in respect of the shopping frequency of books (FREQREBOOK), music (FREQREMUSI) and movies (FREQREMOVI) due to recommendations. In survey AUM, women buy books significantly more often because of recommendations. In survey AON, books are also more often bought by women because of recommendations (mean rank of male respondents is 125.16 vs. 129.09 of female respondents). However, th.is relationship is not significant. On the other side, in survey AON a significant positive relationship between male survey participants and the frequency of buying music by reason of recommendations is detected, which is not the case in survey AUM.



Table 5.13 shows the results of the bi-variate correlation analysis of shopping frequency due to recommendations, time spent on the Internet, age, and educational level of the respondents. The time spent surfing on the Internet has a positive relationship with the shopping frequency of the three product categories except in one case, i.e. in survey AON the shopping frequency does not meet the desired significance level of 0.05 (as shown in the table "Sig." is 0.08). Regarding the age of the respondents, the two surveys only deliver an identical picture in the category movies. Younger people buy movies significantly more often because of recommendations in both surveys. As the table shows, the relationship between educational level and buying frequency of the three products is not well founded. Only in survey AUM, a significant negative relationship between these two factors is identified.

Figure 5.19: Sample description: "I would buy the recommended book instantly at the online--shop that has given a recommendation"

Table 5.14: Regression analysis survey AUM: Factors that influence "I would buy the recommended book instantly at online--shop that has given an interesting recommendation" ~---------rr.--,-T""""..,.,.-,....,.,---..-------, **Adjueted ~td. tirror of** 


The results in Figure 5.19 refer to the question if consumers would buy an interesting book recommended to them in an online shop. Respondents were asked for their level of consent to the hypothetical statement that they would instantly buy a book from an online--shop that has given an interesting book recommendation. This degree of consent is measured on a seven point Likert

Table 5.15: Regression analysis survey AON: Factors that influence "I would buy the recommended book instantly at online-shop that has given an interesting recommendation"


scale ranging from totally disagree to totally agree. The overall tendency in both surveys is that the respondents rather disagree this statement. The arithmetic mean in survey AUM is 3.47 compared to 3.18 in survey AON.

Interestingly, the regression analyses in Table 5.14 and Table 5.15 present that the general impulse buying tendency (IMPUBU) and the positive past experience with online product recommendations (EXRECO) significantly determine the degree of agreement to the statement "I would buy the recommended book instantly at online shop that has given an interesting recommendation". Hence, it can be said that people, who have a high impulse buying tendency and who have a positive past experience with online recommendations are more likely to buy recommended books instantly.


Std. Error

0.38

0.05

0.04

0.04

B

6 31



0.11

(Constant)

TRUSSH

EXSHOP

PRIVCO

Table 5.16: Regression analysis survey AUM: Factors that influence "I would rather

Figure 5.20 shows the results regarding the consent to the statement that they would rather buy a recommended book in a traditional bricks-and-mortar

Beta



0.10

Sir.

0.00

0 00

0.00

16.62


2.81 0.01


Figure 5.20: Sample description: "I would rather buy a recommended book in a bricks-and-mortar store"

Table 5.17: Regression analysis survey AON: Factors that influence "I would rather buy a recommended book in a bricks-and-mortar store"


store. Again the underlying assumption is that they recieve a book recommendation from an online store and that they are interested in that recommended book. In both surveys, the majority adopts a neutral position (AUM 38.5%, AON 37.4%). In survey AUM the proportion of respondents that disagrees (i.e. people that answered "rather disagree", "predominantly disagree",

or "completely disagree") is higher than the people that agree this statement (35.1% disagreement vs. 26.5% agreement). In survey AON, the opposite is the case (23.4% disagreement vs. 39.1% agreement). The arithmetic mean in sw-vey AUM is 3.90 and 4.42 in smvey AON respectively.

The regression analyses in Table 5.16 and Table 5.17 show the factors that influence the degree of consent to the statement "I would rather buy a recommended book in a bricks-and-mortar store" . The author assumed that positive past experience with online shopping (EXSHOP), trust in online shopping (TRUSSH) and privacy concerns (PRIVCO) have an influence. As shown in both surveys both positive past experience with online shopping and trust in online shopping have a significant negative influence on the agreement to statement. In survey AUM, privacy concerns also seem to have an (positive) effect on the degree of acceptance of the statement. However, in survey AON no significant relationship is detected.


Table 5.18: San1ple description: Benefits of recommendations

Figure 5.21 illustrates the degree of consent of the respondents regarding the benefits of recommendations. Respondents were asked to answer this question on a seven point Likert sea.le ranging from totally disagree (1) to totally agree (7). The results refer to respondents that have already received recommendations. Table 5.18 shows that the respondents in both surveys agree that recommendations help to find new interesting products and call attention to low-priced products. They rather disagree that recommendations help to avoid mispurchases and ease navigation in online shops.

Figure 5.21: Sample description: Benefits of recommendations.

Table 5.19: Sample description: Attitudes towards implicit data collection


In Figure 5.22 attitudes towards implicit data collection for generating recommendations are depicted. The survey participants were required to specify their degree of consent to statements regarding implicit data collection (i.e. monitoring user behavior) in online shops. Again, a seven point Likert scale ranging from totally disagree (1) to totally agree (7) is used for measurement.

Figure 5.22: Sample description: Attitudes towards implicit data collection

The statements OBSESU01 to OBSESU03 highlight positive characteristics of implicit data collection, whereas statement OBSESU04 to OBSESU06 reflect critical aspects of monitoring user behavior. In Table 5.19 arithmetic means are shown. In both surveys, the participants rather agree that implicit data monitoring is a suitable method to learn preferences (OBSESU1). Interestingly, the respondents agree that using implicit methods leads to losing control with respect to the estimation of preferences (OBSESU05). However, the participants rather do not agree that this results in a lower recommendation quality (OBESU06).

Further, a bivariate correlation analysis is performed to investigate dependencies between the statements regarding implicit data collection and other factors. Table 5.20 shows statements that refer to the positive aspects of monitor-


Table 5.20: Bi-variate correlation analysis: Attitudes towards implicit data collection

Table 5.21: Mann-Whitney test survey AUM: Gender-specific differences in respect to implicit data collection


ing user behavior have a significant positive relationship with the past positive experience with recommendations (EXRECO) in both surveys a = 0.05). No

Table 5.22: Mann-Whitney test survey AON: Gender-specific. differences in respect to implicit data collection


other significant relationships were detected. Statements that accentuate negative aspects have significant inverse relationships with privacy concerns and a positive relationship with trust in online shopping respectively. Further, people that have experienced positive recommendations in the past (EXRECO) rather reject that user monitoring is questionable from a privacy perspective (OBSESU04). People with a higher educational level a.re more likely to agree that user monitoring results in losing control of the estimation of preferences (OBSESU05) in both surveys. Referring to privacy issues (OBSESU04) and the educational level the situation is ambiguous. Although a significant positive relationship exists in survey AUM, in survey AON a significant relationship is not given on a significance level of 0.05 (Sig. is 0.08). As shown, the age of the respondents does neither influence statements that mention positive aspects of user monitoring nor statements that influence negative aspects. Additionally, a Mann-Whitney test shows no differences between male and female respondents regarding these statements in both surveys.

Figure 5.23 illustrates the imp01tance of explanations for the survey respondents. Explanations expose the reasoning behind the recommendation ( see Section 3.2). As shown, explanations are a fairly important issue for the respondents in both surveys. The arithmetic mean is 4.80 in survey AUM and 4.42 in survey AON respectively.

Figure 5.23: Sample description: Importance of explanations

Table 5.23: Mann-Whitney test survey AUM: Gender-specific differences regarding explanations


Table 5.24: Mann-Whitney test survey AON: Gender-specific differences regarding explanations


Table 5.23 and Table 5.24 depict that gender-specific differences regarding the importance of explanations (FUNCEXPL) are non existent in both surveys. Further, Table 5.25 illustrates the results of a bi-variate correlation analysis. Past positive experience with recommendations (EXRECO) and the importance of explanation are significantly positively interrelated in both surveys (a = 0.05), i.e. respondents with a positive experience are more likely to expect explanatory capabilities from recommender systems. No relationships are detected between educational level, trust in online shopping (TRUSSH) and privacy concerns (PRIVCO). In regard to the age of respondents no univocal results are given. Whereas a significant inverse relationship is found in survey AON, in survey AUM the required significance level of 0.05 is not met (Sig. is 0.07).

Table 5.25: Bi-variate correlation analysis: Importance of explanations


Table 5.26: Sample description: Delivery of recommendations


Figure 5.24 refers to the delivery of online product recommendations. A seven point Likert scale ranging from totally disagree (1) to totally agree (7) is used to measure the degree of consent to specific modes of delivery. As Table 5.26 illustrates, the respondents prefer recommendations on explicit request in both surveys (DELIRE04). Furthermore, respondents rather agree to receive recommendations immediately after logging into the shop (DELIRE01).

Figure 5.24: Sample description: Delivery of recommendations 

The respondents rather reject to receive recommendations by traditional mail (DELIRE07) and while browsing the shop (DELIRE02}. 

In Figure 5.25 assigned motives for provision of recommendations are illustrated. Here, the slU'vey participants were asked to estimate interests e--- vendors pursue with the employment of online product recommendations. A seven point Likert scale ranging from totally disagree (1) to totally agree (7) was used to measure statements regarding the asswned interests of e---vendors. As Table 5.27 shows the highest degree of consent is found regarding the increase of sales (MOTIVE05} in both surveys. Further, the survey participants think that e---vendors employ recommender systems to learn consumer trends (MOTIVE08). The respondents rather disagree that e---vendors try to sell shelf-warmers (MOTIVE02} or customer data (MOTIVE07}. 

Figure 5.25: Sample description: Assigned motives for the provision of recommendations

#### 5.1.5 Ratings and Comments

In this section the importance of product-related reviews in form of ratings and text comments for the buying decision is discussed. In addition, motives for submitting reviews are illustrated.

Figure 5.1.5 refers to the importance of product-related ratings and text comments for the buying decision of the survey participants. The respondents were asked to answer this question on a seven point Likert scale ranging from very unimportant to very important. Sample size, missing values, arithmetic means and rank of importance for both surveys are depicted in Table 5.28. The table


Table 5.27: Sample description: Assigned motives for the provision of recommendations VE05

Table 5.28: Sample description: Importance of different sources of productrelated reviews S RVEY UM RAT! US RATIV ND OMMTHIR OMMVEND


shows that the two surveys show identical results in respect of the importance of the different kinds of reviews (i.e. comments and ratings). Text comments from independent third parties (COMMTHIR) are rated *as* most important for the buying decision, followed by ratings from independent third parties (RATITHIR). Text comments from customers of an online shop (COMM-CUST) are the 3rd most important source of product-related information, followed by ratings from other customers (RATICUST). The least important categories are text comments provided by employees of thee-vendor (COM-MVEND)aud ratings provided by employees of thee-vendor (RATIVEND).

A Whitney-Mann test was performed to investigate gender-specific differences in regard to the importance of the different sources of ratings and comments. As Table 5.29 and Table 5.30 show, in both studies no differences between male and female respondents were detected (et= 0.05).

Figure 5.26: Sample description: Importance of ratings and comments for buying decisions

Table 5.31 summarizes a bi-variate correlation analysis between the different kinds of product-related reviews and past positive experience with online product recommendations (EXRECO), education and age of the respondents. As shown, a significant positive relationship between past positive experience with recommendations and the interest in all the different forms of ratings and comments exists in both surveys. In other words, the more positive experience with recommendations the respondents had in the past, the more they are interested in different forms of comments and ratings. Interestingly, the correlation coefficients illustrate that comments and ratings from other customers are the most valuable source of information for respondents that already have a positive experience with recommendations. Regarding the educational level the correlation analysis shows a significant relationship between comments and ratings from the e-vendor in both surveys. That is, the lower the educational



Table 5.30: Mann-Whitney test survey AON: Gender-specific differences of importance of ratings and comments


level of the respondent the more important are comments and ratings that stem from the e-vendor to him or her. Further, a significant negative relationship between the age and the interest in comments and ratings from other customers exits in both surveys, i.e. the younger the respondents the more likely they are interested in comments and ratings from other customers.

Table 5.31: Bi-variate correlation analysis: Importance of comments and ratings from customers, independent third parties and e-vendors


Figure 5.27: Sample description: Frequency of submitting product-related ratings

In Figure 5.27 the frequency of submitting product-related ratings to online shops is illustrated. The results of the two surveys look fairly similar. The majority in both surveys (38.7% in survey AUM and 40.9% in survey AON) have never submitted a product-related rating to an online store. The arithmetic mean of survey AUM is 2.67 and 2.66 in sw-vey AON respectively.

In respect to the submission of product-related text comments to online shops the frequency of doing this is even lower. As Figure 5.28 shows, in survey AUM 46% of the respondents have never provided comments. In sw-vey AON, 45.9% have never submitted this kind of information to an online shop. The arithmetic mean accounts for 2.30 in survey AON and 2.34 in survey AUM respectively.

Figure 5.28: Sample description: Frequency of submitting product-related text comments

To determine gender specific differences in respect to the submission frequency of ratings and comments two Mann-Whitney tests were performed. As Table 5.32 and Table 5.33 indicate, significant differences whei-e found in both surveys(a = 0.05). In both surveys male respondents submit ratings (FRE-QRATI) as well as comments (FREQCOMM) more often than female respondents.

Table 5.32: Mann-Whitney test survey AUM: Gender-specific differences in respect to the submission of ratings and comments


Table 5.33: Mann-Whitney test survey AON: Gender-specific differences in respect to the submission of ratings and comments


The correlation analysis in Table 5.34 shows interrelationships of the frequency of submitting ratings and comments with other factors (a= 0.05). As shown a significant inverse relationship exists between the frequency of providing ratings as well as comments and the age of the respondents (RESPOAG E) in both surveys. Hence, younger people seem to provide ratings and comments more often. No significant relationships were found between the educational level and the frequency of submission. Furthermore, privacy concerns (PRIVCO) show no significant influence on the submission frequency. Previous positive experience with recommendations (EXRECO) and shopping (EXSHOP) as well as the time spent on the Internet (INTEHOUR) have a significant positive relationship to the submission frequency.

Figure 5.29 refers to motives for submitting product-related reviews of survey participants that have already submitted reviews. The survey participants where asked to specify the degree of consent of statements regarding motives for submitting reviews. A seven point Likert scale ranging from totally disagree

Table 5.34: Bi-variate correlation analysis: Frequency of submitting productrelated ratings and comments


Figure 5.29: Sample description: Motives for submitting reviews.


Table 5.35: Sample description: Motives for submitting reviews

(1) to totally agree (7) was used to measure the degree of acceptance of the respondents. Table 5.35 compares the arithmetic means of the motives for submitting reviews. The higher the arithmetic means the higher is the degree of consent of the survey participants. The results clearly show that getting a better reputation within the virtual community is not a very important motive for the survey participants. To communicate personal tastes and preferences in order to get better personalized recommendations also is not that important for the participants. In contrast, communicating either positive or negative experiences with products is an important motive in both surveys.



In Figure 5.30 and Table 5.36 motives that prevent survey participants from providing ratings and comments are shown. The results refer to participants that have never submitted a rating or comment. The results suggest that having little experience with the product is a very important factor for not

Figure 5.30: Sample description: Motives for not submitting reviews

submitting reviews. The impression that reviews offer no benefit to the individual is also an important factor for not submitting reviews in both surveys. Further, text comments are perceived as relative time-consuming amongst the respondents.

#### 5.2 Verification of the Research Model

In the following sections, the hypotheses and the research model presented in Chapter 4 are verified. A factor analysis, a structural equation model that tests the psychographic hypotheses, and regression models that verify further psychographic and sociodemographic hypotheses are examined.

#### **5.2.1 Exploratory Factor Analysis**

This section deals with the results of the exploratory factor analysis. This analysis is performed to check the validity of the measurnment scales in the first smvey ( AUM). Although a confirma.tive factor analysis is inherently performed in the context of the structmal equation model, the exploratory factor analysis is ma.de for the following three reasons:


As mentioned above in survey AUM, the sample size accounts for 682. 606 cases show no missing values in respect to the variables used in the research model. About 20% (i.e.130) of these cases a.re randomly chosen (using SPSS 12.0.1) to perform the factor analysis and a.re removed from the data.set on which the structural equation model is calculated to avoid fitting the "model to the data" . As a consequence 4 76 cases remained for the calculation of the structural equation model.

In accordance with recent literature regarding factor analysis [Rus02, CO05, CH03, PetO0] principal a.xis factoring was used as the factor extraction procedure. Proma.x with Kaiser normalization was employed used as factor rotation method. Factors with Eigenvalues greater than 1 were extracted.

Ka.iser-Meyer-Olkin measure of sampling adequacy (MSA) accounts for 0. 768. According to this criterion the data is well-suited to perform an exploratory factor analysis. An MSA-Value ~ 0.7 indicates a "pretty good" appropriateness of the data for an exploratory factor analysis [BEPW03].


Table 5.37: Factor analysis AUM: Pattern matrix

In Table 5.37 the loadings of the exploratory factor analysis are shown. ĈIMIŠP refers to the "subjective probability of making a mispurchase" facet of the consumer involvement profiles scale by Laurent and Kapferer |KL86|. CISYMB denotes the "symbolic sign or value attributed by the consumer to the product class" facet of the consumer involvement profiles scale [KL86]. CIHEDO marks the "hedonic value of the product class" according to the consumer involvement profiles scale [KL86]. OPLEAD and OPSEEK indicate the opinion leadership and opinion seeking items as proposed by Flynn, Goldsmith, and Eastman [FGE96]. INREAD, INWRIT, and INRECO are scales designed by the author. INREAD measures the "interest in reading evaluations of books from other customers of an online shop". INWRIT denotes the "interest in writing evaluations of books in an online shop". Finally, INRECO refers to the "interest in obtaining personalized book-recommendations in an online shop".

As depicted in the Table 5.37, the opinion leadership measurement scale loads on two independent factors instead of one as expected. As a consequence, nine factors are extracted (using Eigenvalues greater than 1 as a criterion for factor extraction) instead of eight (as eight meastuement scales are employed). The author assumes that this is due to cultural differences and/or the product class (i.e. books). The scale was originally developed and tested with American students. Table 5.38 shows the items of that scale. OPLEADOl to OPLEAD03 are reverse scaled whereas OPLEAD04 to OPLEAD06 are normally scaled. Probably, OPLEAD0l to OPLEAD03 loads on a different factor, because social desirableness may play an important role in the context of books and due to the negative formulation of the questions. The author has decided to employ OPLEAD04 to OPLEAD06 for the estimation of the model, as factor loadings are generally higher and social desirableness might not play such an important role due to the wording of the questions.

Table 5.38: Opinion leaders by Flynn, Goldsmith, and Eastman


of the scales used for the model estimation actor


Table 5.39 illustrates the scales and corresponding items that are used in the structural equation models described in Section 5.2.2. Cronbach's Alpha is depicted to indicate the reliability of the multi-item measures. Values between 0.8 and 0.9 indicate a "moderate to high level reliability" (DeV96]. Values above 0.9 signify a "high level" of reliability [DeV96]. As shown, all the scales used for the model fall into this two categories.

#### **5.2.2 Psychographic Hypotheses - Structural Equation Model**

In this section, the research model in respect of the psychographic factors is verified. The programm AMOS 5.0 is used for the calculation of the model. Maximum likelihood {ML) is used as method for the estimation of the model para.meters. Maximum likelihood is the most widespread estimation method in international marketing research [HB95b]. This method requires the manifest (i.e. observed) variables to have a multivariate normal distribution. In the context of marketing research, it is very common that the data will fail the assumption of normality. This is also the case in both surveys conducted in the context of this survey. This is assessed by taking a look at Mardia's coefficient and its critical value. In survey AUM Mardia's coefficient accounts for 274.272 and exceeds its critical value of 84.693 by far. Thus, normality of the data can not be assumed. This is also the case in survey AON, where Mardia's coefficient is 169.168 and the critical ratio would be 44.472.

As normality is not given in both surveys the following two options are possible to estimate the model: **(1)** to use an estimation method that does not require a multivariate normal distribution ( e.g. unweighed least squares) or (2) to perform bootstrapping in conjunction with the maximum likelihood estimation method. As maximum likelihood delivers more conservative estimations [AG88] the author has chosen to stick to the maximum likelihood method and to use bootstrapping.

The model consists of 61 variables. 24 variables are observed (i.e. manifest) variables and 37 are unobserved (i.e. latent) variables. 32 of the variables are exogenous and 29 a.re endogenous. In sm-vey AUM, the number of cases is 476. In this survey, originally 606 cases with no missing values regarding the observed variables in the model occurred. However, 130 cases were randomly chosen for the exploratory factor analysis and were not used for the calculation of the model (see Section 5.2.1). In the follow up survey (AON), the number of *cases* with no missing values in respect of the observed variables accounts for 345.

Figure 5.31: Structural equation model: Survey AUM

Figure 5.31 shows the results of the model tested on the dataset of survey AUM. As shown, all regression paths are significant on a level of 0.01. The covariance of the symbolic sign or value and the risk of a mispurchase is not significant.

Figure 5.32 depicts the results of the model with the data from survey AON. Again all regression paths are significant on a significance level of 0.01. The covariance of the hedon.ic value of the product class and the risk of a m.ispurchase is not significant.

The research hypothesis regarding the structural equation model are sununarized in Table 5.40. In both surveys, the hypothesis regarding the model are

Figure 5.32: Structural equation model: Survey AON

of high significance (a: = 0.01). The regression weights are shown to indicate the strength of the influence.

In Table 5.41, the fit indices of the model are depicted. According to the literature, different levels of threshold are depicted for both the GFI and AGFI indices. According to Sharma [Sha96] an AGFI above 0.8 indicates a good model fit. Other researchers (e.g. [HB95a, HATB98]) set the threshold for the AGFI to 0.9. However, both threshold values have no statistical basis. They are derived from practical experience. In literature different fit-indices, their explanatory power and thresholds are lively discussed. For instance, Hu and Bentler generally advise against using the GFI and AGFI as fit indicators [HB99]. In respect to the IFI, TLI, and CFI the established threshold of 0.9 is met. According to Browne and Cudeck a RMSEA below 0.05 indicates a good model fit and a RSMEA below 0.08 indicates a reasonable model



fit (BC93]. The second threshold level is met in both surveys. Hence, the author assumes an adequate model-fit in both surveys.

Table 5.41: Fit-indices of the research model


#### **5.2.3 Psychographic Hypotheses - Regression Model**

In the following, further psychographic determinations of the interest in personalized book recommendations are investigated. These factors include seepticism towards advertising (SKEPAD) [OS98], impulse buying tendency (IM-PUBU) [WJB97], domain-specific innovativeness (DOINNO) [GH91], privacy concerns (PRIVCO), and experience with online shopping (EXSHOP). A regression model was chosen, because including these scales in the structural equation model would lead to a overly complex model.

The measurements where tested in respect to reliability (Cronbach's alpha) and validity (factor analysis with principal axis factoring and promax rotation). As shown in Table 5.42, the scales are well-suited for a regression analysis in terms of reliability and validity.


Table 5.42: Reliability and validity of the scales used for the regression model

Table 5.43: Verification of hypotheses survey AUM: Psychographic factors that influence interest in book recommendations


In Table 5.43 and Table 5.44, the results of the regression model are set out. In both surveys, experience with online shopping and domain specific innovativeness significantly influence the interest in personalized online book recommen-



dations (a = 0.05). According to this results, the hypothesis are summarized in Table 5.45.


Toblo 5.45; Summarization of the nevehographia hymothere

#### 5.2.4 Sociodemographic Hypotheses

This section deals with the verification of the hypotheses that include sociodemographic factors. The question is, whether sociodemographic factors have a significant influence on: (1) the interest in personalized recommendations, (2) the interest in writing book-related reviews, (3) the interest in reading book-related reviews of other consumers.

As shown in Table 5.46, no differences between men and woman are found


Table 5.46: Verification of hypotheses: Gender-specific differences

in both surveys regarding the interest in writing and reading product-related reviews. In survey AON, a gender-specific difference regarding interest in recommendations is found on a significance level of 0.05 but not on a level of 0.01. In survey AUM, a significant difference between male and female respondents is not detected.

Table 5.47 and Table 5.48 investigate, whether educational level and age of the respondents have an influence in regard to the interest in personalized recommendations, in writing book-related reviews, and in reading book-related reviews. As shown, the educational level of the respondents has no significant influence on these three factors. In the contrary, the age of the respondents has a significant influence on the three factors (a = 0.05).

The demographic research hypotheses are summarized in Table 5.49. Besides the standardized regression coefficients the significance level is depicted. "N.S." indicates that on a level of 0.05 the hypothesis is not significant.


Table 5.47: Verification of hypotheses survey AUM: Differences due to age and education


Table 5.48: Verification of hypotheses survey AON: Differences due to age and education


Table 5.49: Summarization of the dem~aphic hypothe.ses

### **Chapter 6**

## **Summary and Directions for Further Research**

The research goal was to examine the underlying psychographic and sociodemographic determinants that influence: (1) the consumer's interest in *personalized recommendations,* (2) the consmner's interest in participating actively in virtual communities of transaction located at online purchase environments by *submitting product-related mtings and comments,* and (3) the consmner's interest in *product-related opinions of other consumers* in virtual communities. This research question was addressed in the context of books that are sold over electronic purchase environments. In the following, the main findings and implications as well as limitations and directions for further research are set out.

#### **6.1 Main Findings**

In the course of the book, two surveys were conducted. The first survey was made in cooperation with the Austrian bookseller A&M Andreas & Dr. Miiller Verlagsbuchhandel (www.aum.at). In this survey the sample size accounts for 682 persons. The second survey was conducted in cooperation with the Internet service provider Telekom Austria AG (www.aon.at). In total 396 respondents filled out the questionnaire of the second slll'vey. The software packages SPSS 12.0.1 and AMOS 5.0 were used for the data analysis.

The main results of the surveys are:


in the e--vendor's online purchase environment given that they are interested in the recommended books.

• Male consumers tend to submit product-related reviews more often than female consumers. Furthermore, younger persons tend to provide reviews more often.

### **6.2 Limitations and Directions for Further Research**

This work is subject to a variety of (partly inevitable) limitations. Firstly, the surveys were posted in German language on web-sited mainly visited by Austrian consumers. Hence, the results may not be representative for consun1ers that stem from other countries or regions. A cross-cultural study would be an interesting point of contact for further research.

Further, using a web-based questionnaire leads to the problem of self-selection of the survey participants. Self-selection refers to the fact that the researcher is not in control of the selection process of the sw-vey participants ( e.g. by selecting paiticipants randomly). For instance, consun1ers with a high interest in the topic may be overrepresented in the sample.

The surveys were limited to a specific type of products. Books were used to test the model. However, the question remains, if the model is still valid with other product classes. For instance, if products are chosen that can be described using "objective criteria" (e.g. personal computers, digital photo cameras) it is questionable, whether opinion leaders are still interested in word-of-mouth from other consuniers and recommendations respectively. Hence, it would be interesting to test the model on other product classes.

The model does not investigate if the interest in recommendations leads to a change in the (buying) behavior of consumers. As the theory of reasoned action (TRA) suggests that a change in the attitude (e.g. the interest in recommendations) is reflected by a change in behavior [AF80], it would be of interest to include behavioral aspects {e.g. adoption of recommendations) in

the model. The technology acceptance model (TAM) [Dav89), which is based on the theory of reasoned action and includes behavioral aspects, should be a promising approach for further research on recommender systems.

The author believes that be has made a relevant contribution to the research regarding online product recommendations. In respect of the growing importance of the field, this work hopefully encourages others to examine further factors that are of relevance for recommendation purposes.

### **Bibliography**


*Americas Conference on Information Systems,* pages 1531-1537, Boston, MA, 2001.



tapestry. *Communications of the ACM,* 35(12):61-70, December 1992.





Downloaded from PubFactory at 01/11/2019 04:40:47AM via free access


#### *BIBLIOGRAPHY*


Downloaded from PubFactory at 01/11/2019 04:40:47AM via free access In Kenneth E. Kendall, editor, *Emerging Information Technologies: Improving decisions, cooperation, and infrastructure,* pages 21-44. Sage, 1999.


## Appendix A

## AMOS Output

### A.1 Survey AUM

Analyala Summary

Date and Time

Date: Oieu..tag, l l. April- 2006 Time: 1:1:26:'l'l

Title

model-aum.bootstrapped: Dien•tag, 11. Aprll 2006 12:26

Groupe

Group number 1 (Group number 1)

Note• ror Group (Group number 1)

The model ia recursive. Sample aise = 476

Variable Summary (Group number 1)

Your model contalna the followlns varlablu (Group number 1)

Obaerved, eado1enou variable, CIHEDO03 CIHEDO02 CIHEDO0I CISYMB03 CISYMB02 CISYMB0I OPLEADO< OPLEAD05 OPLEAD06 INWRIT0I INWRIT02 INWRIT03 INRECO0I INRECO02 INRECO03 INREAD0l INREAD03 INREAD04 OPSEEKOI

OPSEEK04

OPSEEKOS CIMISPO4 CIMISPO3 CIMISPOI Unobserved, endogenous variables OPLEAD INWRIT INRECO INREAD OPSEEK Unobserved, exogenous variables CIHEDO Err3 Err2 Err 1 CISYMB Err6 Err5 Err4 Err10 Err 1 1 Err 12 Res1 Err 16 Err17 Err 18 Res3 Err22 Err23 Err24 Res5 Err19 Err20 Err21 Res4 Err13 Err 1 4 Err15 CIMISP Err9 Err8 Err7 Res2 Variable counts (Group number 1) Number of vari-61 ables in your model: of ob-24 Number served variables: Number of unob-37 served variables: Number of exoge 32 nous variables:
Number of 29 endogenous variables: Parameter summary (Group number 1) Weights Covariances Variances Means 0 0 36 Fixed 1 Labeled o 0 o 0 Unlabeled 25 3 31 0 0 Total 61 3 32 Assessment of normality (Group number 1) Variable min max akew c.r. -0,127 -1.127 CIMISPOI 1 7 CIMISP03 1 7 0,23 2,052 CIMISP04 l 7 0,08 0.536 OPSEEKOS 1 7 0.223 1.982 7 -0,115 -1,023 OPSEEK04 1 OPSEEK01 l T 0,226 2,013 INREAD04 1 7 -0,269 -2,392 INREAD03 1 7 -0.449 -4 1 7 -0.445 -3,962 INREAD01 INRECO03 l 7 -0,131 -1,169

Intercepts

kurtosis











0

o

0

0

Total

37

o

59

છે.

c.r.











Observations farthest from the centrold (Mahalanoble distance) (Group number 1)



#### Models

Default model (Default model)

Notes for Model (Default model)

Computation of degrees of freedom (Default model)

300 Number of distinct sample moments: Number of dis- ସବ tinct parameters to be estimated: Degrees of freedom (300 - 59): 241

Result (Default model)

Minimum was achieved Chi-square = 734,150
Degrees of freedom = 241 Probability level = ,000

Group number 1 (Group number 1 - Default model)

Estimates (Group number 1 - Default model)

Scalar Estimates (Group number 1 - Default model)

Maximum Likelihood Estimates

Regression Weights: (Group number 1 - Default model)



Covariances: (Group number 1 - Default model)

<-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------くー

<—

<--

CIMISPO4

CIMISP03

CIMISP01


0,908

0,89

0,725

0,673

CIMISP

CIMISP

CIMISP

Correlations: (Group number 1 - Default model)


Varlances: (Group number 1 - Default model)


Squared Multiple Correlations: (Group number 1 - Default model)


Modification Indices (Group number 1 - Default model)

Covariances: (Group number 1 - Default model)



**Varlancu1 (Group number 1 - Detaul& model)** 

**Po, Chan1e** 

**M.I.** 

**Ft.ear--lon Welshtaa (Group numt>.r 1 • Default model)** 




#### APPENDIX A. AMOS OUTPUT


Bootstrap (Group number 1 - Default model)

Bootstrap standard errors (Group number 1 - Default model)

Scalar Estimates (Group number 1 - Default model)

Regression Welghts: (Group number 1 - Default model)



#### APPENDIX A. AMOS OUTPUT


Bootstrap Confidence (Group number 1 - Default model)

Blas-corrected percentile method (Group number 1 - Default model)

95% confidence intervals (blas-corrected percentile method)

Scalar Estimates (Group number 1 - Default model)

Regression Welghts: (Group number 1 - Default model)


CIHEDO

CISYMB

CISYMB

CISYMB

OPLEAD

OPLEAD

OPLEAD

INWRIT

INWRIT

INWRIT

INRECO

INRECO

INRECO

INREAD

<-

<--

<--

<---

个个

<--

<==

<---

<---

<-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------

<--

<-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------

0,932

0,91

0,952

0,809

0,766

0,945

0,797

0,909

0.633

0,901

0,951

0,729

0,943

0,937

0,9

0,874

0,923

0,738

0,693

0,908

0,721

0,855

0.521

0,836

0,921

0,627

0,896

0,9

Nicolas Knotzer - 978-3-631-75452-8 Downloaded from PubFactory at 01/11/2019 04:40:47AM via free access

0,959

0,941

0,982

0,873

0,823

0,978

0,856

0,957

0.718

0,86

0,972

0,803

0,973

0,966

0,005

0,005

0,003

0,003

0,007

0,006

0,007

0,003

0,006

0,007

0,007

0,006

0.006

0,004

CIHEDO01

CISYMB03

CISYMB02

CISYMB01

OPLEAD04

OPLEAD05 OPLEAD06

INWRIT01

INWRIT02

INWRIT03 INRECO01

INRECO02

INRECO03

INREAD01


#### APPENDIX A. AMOS OUTPUT


#### Minimisation History (Default model)


Bootstrap (Default model)

Summary of Bootstrap Iterations (Default model)

(Default model)


O bootatrap samples were unused because of a singular covariance matrix.
O bootatrap samples were unused because a solution was not found.
500 uable bootstrap samples were ob

Bootstrap Distributions (Default model)

ML discrepancy (implied vs sample) (Default model)


ML discrepancy (lmplled v• pop) (Default model)


K-L overoptlmlam (unatablllzed) (Default model)


K-L overoptlmlam (atablll11ed) (Default model)


ML dlacrepancy (lmplled va pop) (Default model)




Minimization: 0,031 Miscellaneous: 0,188 2.781 Bootstrap: 3 Total:

#### A.2 Survey AON

Analysis Summary

Date and Time

Date: Dienstag, 11. April 2006 Time: 12:22:39

Title

model aon\_bootstrapped: Dienstag, 11. April 2006 12:22

Groups

Group number 1 (Group number 1)

Notes for Group (Group number 1)

The model is recursive. Sample size = 345

Variable Summary (Group number 1)

Your model contains the following variables (Group number 1)

Observed, endogenous variables CIHEDO03 CIHEDO02 CIHEDO01 CISYMB03 CISYMB02 CISYMB01 OPLEAD04 OPLEAD05 OPLEAD06 INWRITO1 INWRIT02 INWRIT03 INRECO01 INRECO02 INRECO03 INREAD01 INREAD03 INREAD04 OPSEEK01 OPSEEK04 OPSEEK05 CIMISPO CIMISPO3 CIMISPOI Unobserved, endogenous variables OPLEAD INWRIT INRECO INREAD


ables your 24 Number of ob-Number of variables:
Number of unob-37 served variables: 32 Number of exogenous variables:
Number of 29 endogenous variables:

#### Parameter summary (Group number 1)


Assessment of normality (Group number 1)


44.472



#### Models

Default model (Default model)

Notes for Model (Default model)

#### Computation of degrees of freedom (Default model)

Number of dis-300 tinct sample moments: 59 Number of distinct parameters to be estimated: Degrees of freedom (300 - 59): 241

Result (Default model)

Minimum was achieved Chi-square = 696,387
Degrees of freedom = 241 Probability level = ,000

Group number 1 (Group number 1 - Default model)

Estimates (Group number 1 - Default model)

Scalar Estimates (Group number 1 - Default model)

Maximum Likelihood Estimates

Regression Weights: (Group number 1 - Default model)



Standardized Regression Weights: (Group number 1 - Default model)


Covariances: (Group number 1 - Default model)


#### APPENDIX A. AMOS OUTPUT


Squared Multiple Correlations: (Group number 1 - Default model)


Modification Indices (Group number 1 - Default model)

Covariances: (Group number 1 - Default model)



Par
Change

M.I.





Bootstrap (Group number 1 - Default model)

Bootstrap standard errors (Group number 1 - Default model)

Scalar Estimates (Group number 1 - Default model)

Regression Weights: (Group number 1 - Default model)


#### APPENDIX A. AMOS OUTPUT



Bootstrap Confidence (Group number 1 - Default model)

Blas-corrected percentile method (Group number 1 - Default model)

95% confidence intervals (blas-corrected percentile method)

Scalar Estimates (Group number 1 - Default model)

Regression Weights: (Group number 1 - Default model)


#### APPENDIX A. AMOS OUTPUT



Squared Multiple Correlations: (Group number 1 - Delault model)


#### Minimization History (Default model)


Bootstrap (Default model)

Summary of Bootstrap Iterations (Default model)

(Default model)



**0 bootstrap samplee were unuaed becauN of a** •**ingular covariance matrix. 0 bootebap aamplee were unused becauae a solution waa not found. 500 u .. bte bootatrap .. mples were obtained.** 

**Boot**•**trap Dl**•**trlbutlou (Default model)** 

**ML dt.crepancy (lmplled v• aampJe) (Default model)** 


**ML dlacrepancy (Implied v• pop) (Default model)** 


**K-L overoptlmlam (unatablllaed) (Default model)** 



#### K-L overoptimism (stabilized) (Default model)


ML discrepancy (implied vs pop) (Default model)


#### Model Flt Summary



#### Parsimony-Adjusted Measures


#### *APPENDIX A. AMOS OUTPUT*



#### **Execution time aummary**


## Appendix B

### Survey Items


#### **Forschungsergebnlsse der Wlrtschaftsunlversltiit Wien**

Herausgeber: Wirtschaftsunlversitat Wien - vertreten durch a.o. Univ. Prof. Dr. Barbara Sporn 


www.peterlang.de  Michael B. Hinner (ed.)

## The Role of Communication in Business Transactions and Relationships

Frankfurt am Main, Berlin, Bern, Bruxelles, New York, Oxford, Wien, 2007. 462 pp., num. fig.

Intercultural Business Communication. Edited by Michael B. Hinner. Vol. 3 ISBN 978-3-631-54971-9 · pb. € 74.50\*

Without communication, business is not possible. It is, therefore, desirable and necessary that communication be integrated into all aspects of business if one wishes to truly comprehend and succeed in business transactions and relationships. The contributing authors of this volume are all acknowledged experts in the field of communication. Their texts demonstrate how communication influences, directs, and determines virtually each and every face of the business world. In turn, a better, more comprehensive understanding of business is possible.

Contents: M. B. Hinner: Some Thoughts on Perceiving Business Transactions and Relationships · C. R. Berger: Transparent and Opaque Communication · E. T. Hall: Beyond Culture: Context and Meaning · P. Bull: What Is Skilled Interpersonal Communication · A. B. VanGundy: The Care and Framing of Strategic Innovation Challenges · T. G. Plax/L. F. Cecchi: Manager Decisions Based on Communication Facilitated in Focus Groups · D. D. DuFrene/C. Lehman: The Role of Communication in Work Team Success · A. Trethewey/S. Corman: Anticipating K-Commerce: E-Commerce, Knowledge Management, and Organizational Communication · T. A. Avtgis/A. S. Rancer: The Theory of Independent-Mindedness · S. J. Tracy: Becoming a Character for Commerce · Y. Gopal / S. Melkote: New Work Paradigms? · R. Gajjala: Race, Ethnicity, and Intercultural Communication in Online Social Networking and Virtual Work . J. Chunq/G.-M. Cheng: The Relationship between Cultural Context and E-Mail Usage · M. Yzer/M. Fishbein/J. N. Cappella: Using Behavioral Theory to Investigate Routes to Persuasion for Segmented Groups · J. Turow. Audience Construction and Culture Production · O. H. Gandy, Jr. / Z. Li: Framing Comparative Risk

Frankfurt am Main · Bern · Bern · Bruxelles · New York · Oxford · Wien Distribution: Verlag Peter Lang AG Moosstr. 1, CH-2542 Pieterlen Telefax 00 41 (0) 32 / 376 17 27

\*The €-price includes German tax rate Prices are subject to change without notice Homepage http://www.peterlang.de