《电子商务 E-business》阅读文献：Motivating and Supporting User Interaction with Recommender Systems

Motivating and Supporting User Interaction with Recommender Systems Andreas w. neumann Institute of Information Systems and Management Universitat Karlsruhe(TH), 76128 Karlsruhe, Germany a neumann@iism. uni-karlsruhe, de http://www.iism.uni-karlsruhe.de/a.neumann Abstract. This contribution reports on the introduction of explicit rec- mender systems at the University Library of Karlsruhe. In March 2006, a rating service and a review service were added to the already ex- isting behavior-based recommender system. Logged-in users can write re- ews and rate all library documents(books, journals, multimedia, etc. reading reviews and inspecting ratings are open to the general public A role system is implemented that supports the submission of different eviews for the same document from one user to different user groups ( students, scientists, etc. ) Mechanism design problems like bias and fre riding are discussed, to address these problems the introduction of in centive systems is described. Usage statistics are given and the question which recommender system supports which user needs best, is covered Summing up, recommender systems are a way to combine the support of library user interaction with information access beyond catalog searches Keywords: Recommender system, rating service, review service, mech- anism design, incentive system. 1 Introduction The general public is lately becoming accustomed with recommender systems of different kinds at various online stores. But scientific libraries, where the profit contribution of a product (library document) is not the first concern and the costumers(library users) are coming due to very different incentives, are defini- tively a not less promising application area. Due to the supply complexity or the evaluation of the quality, scientists and students are more and more inca- pable of efficiently finding relevant literature in conventional database oriented catalog systems and search engines. A common solution to this problem lies in asking peers(see e. g. [10). Recommender systems aggregate knowledge from many peer groups to the level of expert advice services. They bear the poten ial to significantly reduce transaction costs for literature searches by means of their aggregation capabilities. Scientific libraries are in a good strategic posi- tion to become(even more than now) the information centers of the future 7 Turning library online public access catalogs(OPAC) into customer oriented service portals supporting the interaction of the customers is one step to this L Kovacs. N. Fuhr, and C. Meghini(Eds ) ECDL 2007. LNCS 4675. pp. 428-139, 2007 ringer-Verlag Berlin Heidelberg 200

Motivating and Supporting User Interaction with Recommender Systems Andreas W. Neumann Institute of Information Systems and Management, Universit¨at Karlsruhe (TH), 76128 Karlsruhe, Germany a.neumann@iism.uni-karlsruhe.de http://www.iism.uni-karlsruhe.de/a.neumann Abstract. This contribution reports on the introduction of explicit rec￾ommender systems at the University Library of Karlsruhe. In March 2006, a rating service and a review service were added to the already ex￾isting behavior-based recommender system. Logged-in users can write re￾views and rate all library documents (books, journals, multimedia, etc.); reading reviews and inspecting ratings are open to the general public. A role system is implemented that supports the submission of different reviews for the same document from one user to different user groups (students, scientists, etc.). Mechanism design problems like bias and free riding are discussed, to address these problems the introduction of in￾centive systems is described. Usage statistics are given and the question, which recommender system supports which user needs best, is covered. Summing up, recommender systems are a way to combine the support of library user interaction with information access beyond catalog searches. Keywords: Recommender system, rating service, review service, mech￾anism design, incentive system. 1 Introduction The general public is lately becoming accustomed with recommender systems of different kinds at various online stores. But scientific libraries, where the profit contribution of a product (library document) is not the first concern and the costumers (library users) are coming due to very different incentives, are defini￾tively a not less promising application area. Due to the supply complexity or the evaluation of the quality, scientists and students are more and more inca￾pable of efficiently finding relevant literature in conventional database oriented catalog systems and search engines. A common solution to this problem lies in asking peers (see e. g. [10]). Recommender systems aggregate knowledge from many peer groups to the level of expert advice services. They bear the poten￾tial to significantly reduce transaction costs for literature searches by means of their aggregation capabilities. Scientific libraries are in a good strategic posi￾tion to become (even more than now) the information centers of the future [7]. Turning library online public access catalogs (OPAC) into customer oriented service portals supporting the interaction of the customers is one step to this L. Kov´acs, N. Fuhr, and C. Meghini (Eds.): ECDL 2007, LNCS 4675, pp. 428–439, 2007. c Springer-Verlag Berlin Heidelberg 2007

Motivating and Supporting User Interaction with Recommender Systems 429 goal. Valid and credible information is a scarce resource[ 20. Information con- sumes the attention of its recipients. Hence a wealth of information creates a poverty of attention and a need to allocate that attention efficiently among the overabundance of information sources that might consume it. "211 The more general term"recommender system"was coined by Resnick and Varian to better describe the action than the more narrow " collaborative fil- tering"16. A recommender system reads observed user behavior or opinions from users as input, then aggregates and directs the resulting recommendations to appropriate recipients. Recommender systems can be classified into two dif- ferent main categories. An implicit recommender system is based on behavioral usage data like purchases, digital library catalog inspections, or lending data An explicit recommender system directly asks the users for their opinions on certain objects. A more technical classification with a focus on applications in e-commerce can be found in [18 and [ 19. For a more up-to-date overview on recommender systems e. g. see Adomavicius and Tuzhilin [1]. In Geyer-Schulz et al. 5 an early application of recommender systems including group-specific services in e-learning is presented. Herlocker et al. 9 deals with the technical evaluation of recommender systems The focus of this paper lies on the experiences with motivation and support of interaction between library users at the University Library of Karlsruhe. First the introduced recommender systems are described, then mechanism design is discussed to address motivational problems. Finally, general lessons learned from integrating different recommender systems into large existing legacy library ap- plications are summarized and the evaluation of such systems is discussed All in this paper presented recommender systems are fully operational services accessible by the general public. For further information on how to use these see "parTicipate!"athttp://reckvk.em.uni-karlsruhe.de/.Inanswertostrong privacy concerns among students and scientists all portrayed recommender ser- vices are object-centered. They do not classify the users by observation or asking them for their interest, but they classify and gather data on the documents of a library. Figure 1 shows a cutout of the detailed document inspection page of [13] in the OPac of the University Library of Karlsruhe. The behavior-based ser vice is accessibly by clicking on "Empfehlungen"(Recommendations), the rating service by "Bewertung abgeben"(Submit rating) or direct inspection of"Bew ertung des Titels nach Nutzergruppen"(Ratings of the titles by user group) and finally the review service by "Rezension schreiben"(Write review),"Rezen- sionen anzeigen"(Inspect reviews), and"Meine Rezensionen"(My reviews). All systems are programmed in Perl or PHP(or a combination of both), use Post- greSQL databases, and are running on Linux servers 2 Behavior-Based Recommender service Behavior-based recommender services are observing the behavior of users and thereby implicitly collecting information about the objects the users are inspect g. The necessary homogeneity of a group of users in this case is granted by

Motivating and Supporting User Interaction with Recommender Systems 429 goal. Valid and credible information is a scarce resource [20]. Information con￾sumes the attention of its recipients. “Hence a wealth of information creates a poverty of attention and a need to allocate that attention efficiently among the overabundance of information sources that might consume it.” [21] The more general term “recommender system” was coined by Resnick and Varian to better describe the action than the more narrow “collaborative fil￾tering” [16]. A recommender system reads observed user behavior or opinions from users as input, then aggregates and directs the resulting recommendations to appropriate recipients. Recommender systems can be classified into two dif￾ferent main categories. An implicit recommender system is based on behavioral usage data like purchases, digital library catalog inspections, or lending data. An explicit recommender system directly asks the users for their opinions on certain objects. A more technical classification with a focus on applications in e-commerce can be found in [18] and [19]. For a more up-to-date overview on recommender systems e. g. see Adomavicius and Tuzhilin [1]. In Geyer-Schulz et al. [5] an early application of recommender systems including group-specific services in e-learning is presented. Herlocker et al. [9] deals with the technical evaluation of recommender systems. The focus of this paper lies on the experiences with motivation and support of interaction between library users at the University Library of Karlsruhe. First, the introduced recommender systems are described, then mechanism design is discussed to address motivational problems. Finally, general lessons learned from integrating different recommender systems into large existing legacy library ap￾plications are summarized and the evaluation of such systems is discussed. All in this paper presented recommender systems are fully operational services accessible by the general public. For further information on how to use these see “Participate!” at http://reckvk.em.uni-karlsruhe.de/. In answer to strong privacy concerns among students and scientists all portrayed recommender ser￾vices are object-centered. They do not classify the users by observation or asking them for their interest, but they classify and gather data on the documents of a library. Figure 1 shows a cutout of the detailed document inspection page of [13] in the OPAC of the University Library of Karlsruhe. The behavior-based ser￾vice is accessibly by clicking on “Empfehlungen” (Recommendations), the rating service by “Bewertung abgeben” (Submit rating) or direct inspection of “Bew￾ertung des Titels nach Nutzergruppen” (Ratings of the titles by user group), and finally the review service by “Rezension schreiben” (Write review), “Rezen￾sionen anzeigen” (Inspect reviews), and “Meine Rezensionen” (My reviews). All systems are programmed in Perl or PHP (or a combination of both), use Post￾greSQL databases, and are running on Linux servers. 2 Behavior-Based Recommender Service Behavior-based recommender services are observing the behavior of users and thereby implicitly collecting information about the objects the users are inspect￾ing. The necessary homogeneity of a group of users in this case is granted by

A.W. Neumann Rezension schreiben Bewertung abgeben Rezensionen anzeige Meine Rezensionen Empfehlungen Bewertung des Titels nach 食★★★★ Studenten:35(4Bew) ★☆★★ o Mtarbeiter:445Bew Fig 1. Recommender start interface on a document's detailed inspection page sion schreiben- Write review: Bewertung abgeben- Submit rating: Rezensionen nzeigen-Inspect reviews: Meine Rezensionen- My reviews: Empfehlungen mendations: Bewertung des Titels nach Nutzergruppen- Ratings of the titles by user group the principle of self-selection [17, 22 In a library setting usage behavior can be observed at different stages: detailed inspections of documents in the OPAc ordering paper documents from the magazine, ordering paper documents that are currently lent, and finally picking-up a paper document or downloading a file from the digital library. The main concern for the data selection is bias. It can be shown that lending and ordering data is highly biased, since e.g. stu- dents very often do not order the book they are mostly interested in, because most likely it is already lent, but actually their consideration-set only includes documents that they will be able to obtain timely before the corresponding ex amination. In marketing several conceptual models which describe a sequence of sets(e.g. total set 2 awareness set 2 consideration set 2 choice set(11 p. 153)) have been developed to describe such situations [14, 23. For this reason the behavior-based recommender service at the University Library of Karlsruhe based on anonymized OPAC searches(hits on document inspection pages)and ot on lending data. Due to transaction costs the detailed inspection of docu- ments in the OPAC of a library can be put on a par with a purchase incidence in a consumer store setting. A market basket consists of all documents that have been co-inspected by one anonymous user within one session. To answer the question, which co-inspections occur non-random, an algorithm based on calculating inspection frequency distribution functions following a logarithmic series distribution(LSD)is applied 6. Such a recommender system is opera- tional at the OPac of the University Library of Karlsruhe in a first version since

430 A.W. Neumann Fig. 1. Recommender start interface on a document’s detailed inspection page. Rezen￾sion schreiben – Write review; Bewertung abgeben – Submit rating; Rezensionen anzeigen – Inspect reviews; Meine Rezensionen – My reviews; Empfehlungen – Recom￾mendations; Bewertung des Titels nach Nutzergruppen – Ratings of the titles by user group. the principle of self-selection [17,22]. In a library setting usage behavior can be observed at different stages: detailed inspections of documents in the OPAC, ordering paper documents from the magazine, ordering paper documents that are currently lent, and finally picking-up a paper document or downloading a file from the digital library. The main concern for the data selection is bias. It can be shown that lending and ordering data is highly biased, since e.g. stu￾dents very often do not order the book they are mostly interested in, because most likely it is already lent, but actually their consideration-set only includes documents that they will be able to obtain timely before the corresponding ex￾amination. In marketing several conceptual models which describe a sequence of sets (e. g. total set ⊇ awareness set ⊇ consideration set ⊇ choice set ([11], p. 153)) have been developed to describe such situations [14,23]. For this reason the behavior-based recommender service at the University Library of Karlsruhe is based on anonymized OPAC searches (hits on document inspection pages) and not on lending data. Due to transaction costs the detailed inspection of docu￾ments in the OPAC of a library can be put on a par with a purchase incidence in a consumer store setting. A market basket consists of all documents that have been co-inspected by one anonymous user within one session. To answer the question, which co-inspections occur non-random, an algorithm based on calculating inspection frequency distribution functions following a logarithmic series distribution (LSD) is applied [6]. Such a recommender system is opera￾tional at the OPAC of the University Library of Karlsruhe in a first version since

Motivating and Supporting User Interaction with Recommender Systems 431 nmen mit folgenden Titeln auger 2006 1. Hgh Performance Linux Clusters / Sloan, Joseph D, 2005. (16) 围园 verteilte Programmierung /Rauber, Thomas: RUnger, Gudula, 2000. (10)a E d parallel computing 4. Using MPl/ Gropp, D: Lusk, Ewing L; Skiellum, Anthony, 1999, (9) 7) 回回回回回回 9. Beowulf cluster computing with Linux/ Sterling, Thamas Lawrence, 2002.(6 11. Custern mit Hintergrundwssen /Hotho, Andreas, 2004,(6) 2. C und Linux/ Grafe, Martin, 2005. (3) Fig. 2. Recommendation list of"Cluster computing" by Bauke and Mertens. The num- ber of co-inspections is given in brackets after each title June 2002 8 and in the current web service version(facilitating WSDL, XML and SOAP) since January 2006 Figure 2 shows the recommendation list of"Cluster computing"by Bauke nd Mertens(cut-out from the web page). The number of co-inspections is given in brackets after each title. Documents with less than three co-inspections have been rated by the lsd test to be not significantly related to this book. Since the usage distribution of documents in nearly every library is highly skewed(newer documents, or documents to topics that interest a large part of the overall library users, in general are more requested), many recommendations will be generated for documents that are used often while seldom used documents have fewer or no recommendations. Recommendations are updated daily. Of the 929 637 doc uments in the catalog, 192 647 documents have lists with recommendations, a total of 2 843017 recommendations exist. Because of the skewness, the coverage of actual detailed document inspections is 74.9%(much higher than the cover age of the complete catalog). So the probability that recommendations exist for a document a user is currently interested in is 0.749(status of 2007-03-19).A user survey asking the library users"I consider the recommendation service in general "on a Likert scale from 1(very bad) to 5(very good) yielded a mean of 4. 1 from 484 votes between 2005-03-21 and 2006-03-06. This type of recom- mender service is best suited to users trying to find standard literature or further standard readings of a field corresponding to the document they are currently nspecting. Although it does not support the direct interaction(communication) between customers, everybody using the service profits from the actions of other library users An e-mail notification service was added at a later stage. Users with a library account receive an e-mail including a direct link to the recommendation page if

Motivating and Supporting User Interaction with Recommender Systems 431 Fig. 2. Recommendation list of “Cluster computing” by Bauke and Mertens. The num￾ber of co-inspections is given in brackets after each title. June 2002 [8] and in the current web service version (facilitating WSDL, XML and SOAP) since January 2006. Figure 2 shows the recommendation list of “Cluster computing” by Bauke and Mertens (cut-out from the web page). The number of co-inspections is given in brackets after each title. Documents with less than three co-inspections have been rated by the LSD test to be not significantly related to this book. Since the usage distribution of documents in nearly every library is highly skewed (newer documents, or documents to topics that interest a large part of the overall library users, in general are more requested), many recommendations will be generated for documents that are used often while seldom used documents have fewer or no recommendations. Recommendations are updated daily. Of the 929637 doc￾uments in the catalog, 192647 documents have lists with recommendations, a total of 2 843017 recommendations exist. Because of the skewness, the coverage of actual detailed document inspections is 74.9% (much higher than the cover￾age of the complete catalog). So the probability that recommendations exist for a document a user is currently interested in is 0.749 (status of 2007-03-19). A user survey asking the library users “I consider the recommendation service in general” on a Likert scale from 1 (very bad) to 5 (very good) yielded a mean of 4.1 from 484 votes between 2005-03-21 and 2006-03-06. This type of recom￾mender service is best suited to users trying to find standard literature or further standard readings of a field corresponding to the document they are currently inspecting. Although it does not support the direct interaction (communication) between customers, everybody using the service profits from the actions of other library users. An e-mail notification service was added at a later stage. Users with a library account receive an e-mail including a direct link to the recommendation page if

432 A.W. Neumann new recommendations appear for a previously specified document. The usage of this service didn't meet the first expectations. Users seem to be skeptic about any service that tries to grab their attention(like spam mails)at times when they didn't even visit the library. To overcome this problem it is planned to extend this notification service in the near future to support RSS feed techniques. Thereby, each user can decide within the RSs reader when to poll the service. Further on, this way it is no longer connected to existing user accounts, but opened personalized service to the general public 3 Explicit Recommender Systems Two different kinds of explicit recommender systems are online at the University Library of Karlsruhe since March 2006, a rating service and a review service. To prevent fraudulent use, submitting ratings and reviews is possible only for logged- in users. These services differ from most other systems(e. g. Amazon coms) by means of user and target groups and strict separation of ratings and reviews Currently three different user groups exist: students(Studenten), university staff Mitarbeiter), and others(Externe)not directly associated with the universit While one could easily come up with more elaborate user classifications, these vices. They are checked(and afterwards tracked over time)by th ag these ser- three groups have been used by the library for many years prece e library for each user before handing out the library card. The guarantee of correctness made this user classification the(pragmatic) choice of approach for a library with an existing base of approximately 24 400 active users 3.1 Rating Service This service allows logged-in users so submit a numerical rating for a document on a Likert scale from 1 (very bad) to 5(very good). Every user can submit only one rating per document. The ratings are aggregated for each user group separately and are shown in numerical form(average rating, number of ratings) as well as an enlightened-star-graphic on the detailed document inspection page In figure I we see 4 ratings from students(Studenten) with an average of 3.5 and 5 ratings from university staff(Mitarbeiter) with an average 4.4. Thus, at a first glance[13 seems to be an overall good book, even more praised by scientists than by students Figure 3 shows the overall number of ratings online from 2006-03-03 to 2007 03-19. One large draw back of the current setup is known. Users searching the catalog are normally not logged-in until they want to order a paper copy of a document not freely available right now. To submit ratings they have to first log-in. This hurdle seems to have a huge influence on the number or submitted ratings, although it should have a very positive influence on the quality of the ratings. This service is best suited to get a first quick estimation of the overall quality of a document within certain user groups

432 A.W. Neumann new recommendations appear for a previously specified document. The usage of this service didn’t meet the first expectations. Users seem to be skeptic about any service that tries to grab their attention (like spam mails) at times when they didn’t even visit the library. To overcome this problem it is planned to extend this notification service in the near future to support RSS feed techniques. Thereby, each user can decide within the RSS reader when to poll the service. Further on, this way it is no longer connected to existing user accounts, but opened as a personalized service to the general public 3 Explicit Recommender Systems Two different kinds of explicit recommender systems are online at the University Library of Karlsruhe since March 2006, a rating service and a review service. To prevent fraudulent use, submitting ratings and reviews is possible only for logged￾in users. These services differ from most other systems (e. g. Amazon.com’s) by means of user and target groups and strict separation of ratings and reviews. Currently three different user groups exist: students (Studenten), university staff (Mitarbeiter), and others (Externe) not directly associated with the university. While one could easily come up with more elaborate user classifications, these three groups have been used by the library for many years preceding these ser￾vices. They are checked (and afterwards tracked over time) by the library for each user before handing out the library card. The guarantee of correctness made this user classification the (pragmatic) choice of approach for a library with an existing base of approximately 24400 active users. 3.1 Rating Service This service allows logged-in users so submit a numerical rating for a document on a Likert scale from 1 (very bad) to 5 (very good). Every user can submit only one rating per document. The ratings are aggregated for each user group separately and are shown in numerical form (average rating, number of ratings) as well as an enlightened-star-graphic on the detailed document inspection page. In figure 1 we see 4 ratings from students (Studenten) with an average of 3.5 and 5 ratings from university staff (Mitarbeiter) with an average 4.4. Thus, at a first glance [13] seems to be an overall good book, even more praised by scientists than by students. Figure 3 shows the overall number of ratings online from 2006-03-03 to 2007- 03-19. One large drawback of the current setup is known. Users searching the catalog are normally not logged-in until they want to order a paper copy of a document not freely available right now. To submit ratings they have to first log-in. This hurdle seems to have a huge influence on the number or submitted ratings, although it should have a very positive influence on the quality of the ratings. This service is best suited to get a first quick estimation of the overall quality of a document within certain user groups