It can be seen from Table 2 that the proposed gsb algorithm outperforms other algorithms significantly on both datasets, which demonstrates the effectiveness our approach. According to t-test, GSB is significantly better than the second best algorithm GTB, with p<. 001 on all evaluation metrics for both datasets. Despite its simplicity, tag generalization has led to quite sizable improvements on both the subject-based and topic-based algorithms, confirming our hypothesis about the reliability of matrix UT and IT. The fusion method focuses on occurrences of tags rather than their weighing values, thus are not able to benefit from tag generalization. The subject-based algorithms, with or without tag generalization, have shown superiority over the topic-based algorithms, demonstrating the advantages of extracting subjects for recommendation. In general, the tag-enhanced algorithms outperform the traditional CF approaches strikingly, especially on the sparser Large dataset. It implies that it is of great potential to improve item recommendation quality leveraging tagging information, especially when the dataset is sparse. The prediction accuracy of the implemented algorithms on the Small dataset is much higher compared with that on the Large dataset, which may due to the fact that the Small dataset is much denser than the large dataset and more data are available for training/profiling 5 Conclusions and future research This paper discusses a subject-centered model of collaborative tagging, in which subjects are assumed to play the key role in determining users' tagging behavior. We propose the concept of Consistent Nonnegative Matrix Factorization and use it to discover the hidden subjects and inner relations within this model from tagging data. We also propose a preprocessing technique called tag generalization to remove noises stemming from meaningless tags. An evaluation study using two real-world collaborative tagging datasets from Delicious has demonstrated the effectiveness of our approach. For future work, we plan to design more sophisticated techniques for tag generalization, and apply it to matrix IT as well. We will also investigate how to optimize various control parameters of our approach(e. g, c in Equation 1, k, the target number of subjects to be clustered) automatically Acknowledgements The authors wish to acknowledge research support from the NNSFC (60621001, 60875049, 70890084), MOST (2006AA010106), and CAs(2F07C01, 2F08N03) References Halpin, H, Robu, V, and Shepherd, H. "The complex dynamics of collaborative tagging, " in Proceedings of t he 16th i nternational conference on W orld w ide Web, ACM, Banff, Alberta Lambiotte, R, and Ausloos, M. "Collaborative Tagging as a Tripartite Network, in: Lecture Notes in Computer Science In ICCS 2006, 2006, pp. 1114-1117 Lee, DD, and Seung, H.S. "Learning the parts of objects by non-negative matrix factorization, "Nature (401:6755)1999pp788-791 Peng, J, and Zeng, D. Topic-based Web Page Recommendation Using Tags, "in: Proceedings of The 2nd International Workshop on Social Computing, IEEE, Dallas, Texas, 2009 Tso-Sutter, K H.L., Marinho, L B, and Schmidt-Thieme, L. Tag-aware recommender systems by fusion f collaborative filtering algorithms, in: Proceedings of t he A CM s ymposium on A pplied computing, ACM, Fortaleza, Ceara, Brazil, 2008 Xu, W, Liu, X, and Gong, Y "Document clustering based on non-negative matrix factorization, Proceedings oft he 26t h annu al i nternational ACM SI GIR c onference onR esearch and levelopment in informaion retrieval, ACM, Toronto, Canada, 2003 Zhao, S, Du, N, Nauerz, A, Zhang, X, Yuan, Q, and Fu,R "Improved recommendation based on collaborative tagging behaviors, "in: Proceedings of the 13t h i nternational conference on Intelligent user interfaces, ACM, Gran Canaria, Spain, 2008 19th Workshop on Information Technologies and Systems
It can be seen from Table 2 that the proposed GSB algorithm outperforms other algorithms significantly on both datasets, which demonstrates the effectiveness our approach. According to t-test, GSB is significantly better than the second best algorithm GTB, with p<0.001 on all evaluation metrics for both datasets. Despite its simplicity, tag generalization has led to quite sizable improvements on both the subject-based and topic-based algorithms, confirming our hypothesis about the reliability of matrix UT and IT. The fusion method focuses on occurrences of tags rather than their weighing values, thus are not able to benefit from tag generalization. The subject-based algorithms, with or without tag generalization, have shown superiority over the topic-based algorithms, demonstrating the advantages of extracting subjects for recommendation. In general, the tag-enhanced algorithms outperform the traditional CF approaches strikingly, especially on the sparser Large dataset. It implies that it is of great potential to improve item recommendation quality leveraging tagging information, especially when the dataset is sparse. The prediction accuracy of the implemented algorithms on the Small dataset is much higher compared with that on the Large dataset, which may due to the fact that the Small dataset is much denser than the Large dataset and more data are available for training/profiling. 5. Conclusions and Future Research This paper discusses a subject-centered model of collaborative tagging, in which subjects are assumed to play the key role in determining users’ tagging behavior. We propose the concept of Consistent Nonnegative Matrix Factorization and use it to discover the hidden subjects and inner relations within this model from tagging data. We also propose a preprocessing technique called tag generalization to remove noises stemming from meaningless tags. An evaluation study using two real-world collaborative tagging datasets from Delicious has demonstrated the effectiveness of our approach. For future work, we plan to design more sophisticated techniques for tag generalization, and apply it to matrix IT as well. We will also investigate how to optimize various control parameters of our approach (e.g., c in Equation 1, k, the target number of subjects to be clustered) automatically. Acknowledgements The authors wish to acknowledge research support from the NNSFC (60621001, 60875049, 70890084), MOST (2006AA010106), and CAS (2F07C01, 2F08N03). References Halpin, H., Robu, V., and Shepherd, H. "The complex dynamics of collaborative tagging," in: Proceedings of t he 16th i nternational conference on W orld W ide W eb, ACM, Banff, Alberta, Canada, 2007. Lambiotte, R., and Ausloos, M. "Collaborative Tagging as a Tripartite Network," in: Lecture Notes in Computer Science In ICCS 2006, 2006, pp. 1114-1117. Lee, D.D., and Seung, H.S. "Learning the parts of objects by non-negative matrix factorization," Nature (401:6755) 1999, pp 788-791. Peng, J., and Zeng, D. "Topic-based Web Page Recommendation Using Tags," in: Proceedings of The 2nd International Workshop on Social Computing, IEEE, Dallas, Texas, 2009. Tso-Sutter, K.H.L., Marinho, L.B., and Schmidt-Thieme, L. "Tag-aware recommender systems by fusion of collaborative filtering algorithms," in: Proceedings of t he A CM s ymposium on A pplied computing, ACM, Fortaleza, Ceara, Brazil, 2008. Xu, W., Liu, X., and Gong, Y. "Document clustering based on non-negative matrix factorization," in: Proceedings of t he 26t h annu al i nternational ACM SI GIR c onference on R esearch and development in informaion retrieval, ACM, Toronto, Canada, 2003. Zhao, S., Du, N., Nauerz, A., Zhang, X., Yuan, Q., and Fu, R. "Improved recommendation based on collaborative tagging behaviors," in: Proceedings of the 13t h i nternational conference on Intelligent user interfaces, ACM, Gran Canaria, Spain, 2008. 78 19th Workshop on Information Technologies and Systems
TOWARD MORE DIVERSE RECOMMENDATIONS ITEM RE-RANKING METHODS FOR RECOMMENDER SYSTEMS Gediminas adomavicius YoungOk Kwon Department of Information and Decision Sciences Carlson School of Management, University of Minnesota gedas(@umn.edu, kwonx052@umn.edu Abstract Recommender systems are becoming increasingly important to individual users and businesses. for providing personalized recommendations. However, while the majority of algorithms proposed in recommender systems litera ture have fo cused on imp roving reco mmendation accuracy (a s exemplified by the rece nt Netflix Prize competition), other important aspects of rec ommendation quality, such as the diversity of recommendations, have often been overlooked. In this paper, we introduce a nu mber of item re-ran king methods tha t can g enerate sub stantially more d iverse recommendations across all users w hile m aintaining comparable levels of rec ommendation accuracy. Empi rical resul ts con sistently show t he di versity gai ns oft he proposed re-ranking methods for several real-world rating datasets and different rating prediction techniques der systems, collaborative filtering, recommendation diversity, ranking functions 1. Introduction and motivation In recent years, recommender systems have become an important research topic in academia and industry (Adomavicius and Tuzhilin 2005). However, in most cases, new techniques are designed to improve the accuracy of recommendations, including the most recent algorithms from Netflix Prize competition (netflixprize. com); other aspects, such as the diversity of recommendations, have often been overlooked in evaluating the recommendation quality. The importance of diverse recommendations has been emphasized in several recent studies (Adomavicius and Kwon 2008, Bradley and Smyth 2001 Brynjolfsson et al. 2007, Fleder and Hosanagar 2009, Zhang and Hurley 2008, Ziegler et al. 2005). These studies argue that one of the goals of recommender systems is to provide a user with highly idiosyn or personalized items, and more diverse recommendations result in more opportunities for users to get recommended such items. With this motivation, some studies proposed new recommendation methods that can increase the diversity of recommendation sets for a given individual user(i.e, individual diversity), often measured by the average dissimilarity between all pairs of recommended items( bradley and Smyth 2001, Zhang and Hurley 2008, Ziegler et al. 2005) More diverse recommendations could be beneficial for some businesses as well Brynjolfsson et al. 2007, Fleder and Hosanagar 2009). For example, it would be profitable to Netflix if their recommender system can encourage users to rent more"long-tail" type of movies (i.e, more obscure items that are located in the tail of the sales distribution) because they are typically less costly to license and acquire from distributors than new-release or highly-popular movies of big studios(Goldstein and Goldstein 2006) Few recent studies(Brynjolfsson et al. 2007, Fleder and Hosanagar 2009)started examining the impact of recommender systems on sales diversity by Table 1. Accuracy-diversity tradeoff: example recommendations across all users which will be Quality Metric AccuracyDiversity the focus of this individual diversity of recommendations does Popular Item (item with the largest 49 distinct not necessarily imply high aggregate diversity of known ratings) Items For example, while recommending to all users"Long-Tail"Item(item with the 695 distinct the same five best-selling items that are not smallest number of known ratings) similar to each other will result in high individual diversity, the aggregate diversity will Note, Recommendations for 2828 users by a standard item- based collaborative filtering technique on MovieLens data 19th Workshop on Information Technologies and Systems
TOWARD MORE DIVERSE RECOMMENDATIONS: ITEM RE-RANKING METHODS FOR RECOMMENDER SYSTEMS Gediminas Adomavicius YoungOk Kwon Department of Information and Decision Sciences Carlson School of Management, University of Minnesota gedas@umn.edu, kwonx052@umn.edu Abstract Recommender systems are becoming increasingly important to individual users and businesses for providing personalized recommendations. However, while the majority of algorithms proposed in recommender systems litera ture have fo cused on imp roving reco mmendation accuracy (a s exemplified by the rece nt Netflix Prize competition) , other important aspects of rec ommendation quality, such as the diversity of recommendations, have often been overlooked. In this paper, we introduce a nu mber of item re-ran king methods tha t can g enerate sub stantially more d iverse recommendations across all users w hile m aintaining comparable l evels of rec ommendation accuracy. Empi rical resul ts con sistently show t he di versity gai ns of t he pr oposed re-ranking methods for several real-world rating datasets and different rating prediction techniques. Keywords: recommender systems, collaborative filtering, recommendation diversity, ranking functions. 1. Introduction and Motivation In recent years, recommender systems have become an important research topic in academia and industry (Adomavicius and Tuzhilin 2005). However, in most cases, new techniques are designed to improve the accuracy of recommendations, including the most recent algorithms from Netflix Prize competition (netflixprize.com); other aspects, such as the diversity of recommendations, have often been overlooked in evaluating the recommendation quality. The importance of diverse recommendations has been emphasized in several recent studies (Adomavicius and Kwon 2008, Bradley and Smyth 2001, Brynjolfsson et al. 2007, Fleder and Hosanagar 2009, Zhang and Hurley 2008, Ziegler et al. 2005). These studies argue that one of the goals of recommender systems is to provide a user with highly idiosyncratic or personalized items, and more diverse recommendations result in more opportunities for users to get recommended such items. With this motivation, some studies proposed new recommendation methods that can increase the diversity of recommendation sets for a given individual user (i.e., individual diversity), often measured by the average dissimilarity between all pairs of recommended items (Bradley and Smyth 2001, Zhang and Hurley 2008, Ziegler et al. 2005). More diverse recommendations could be beneficial for some businesses as well (Brynjolfsson et al. 2007, Fleder and Hosanagar 2009). For example, it would be profitable to Netflix if their recommender system can encourage users to rent more “long-tail” type of movies (i.e., more obscure items that are located in the tail of the sales distribution) because they are typically less costly to license and acquire from distributors than new-release or highly-popular movies of big studios (Goldstein and Goldstein 2006). Few recent studies (Brynjolfsson et al. 2007, Fleder and Hosanagar 2009) started examining the impact of recommender systems on sales diversity by considering aggregate diversity of recommendations across all users, which will be the focus of this paper. Note that high individual diversity of recommendations does not necessarily imply high aggregate diversity. For example, while recommending to all users the same five best-selling items that are not similar to each other will result in high individual diversity, the aggregate diversity will Table 1. Accuracy-diversity tradeoff: example Quality Metric: Top-1 recommendation of: Accuracy Diversity Popular Item (item with the largest number of known ratings) 82% 49 distinct items “Long-Tail” Item (item with the smallest number of known ratings) 68% 695 distinct items Note. Recommendations for 2828 users by a standard itembased collaborative filtering technique on MovieLens data. 79 19th Workshop on Information Technologies and Systems