Tags Meet Ratings: Improving Collaborative Filtering with Tag-Based Neighborhood Method Zhe Wang Yo w Hu wu National engineerin National engineerin National Engineering Research Center of Research Center of Fundamental software Fundamental Software Fundamental Software Institute of software Institute of software Institute of software Chinese academy of Chinese academy of Chinese academy of Sciences Sciences Sciences Graduate University of yang@techs. iscas ac cn wuhu@@techs.iscasaccn Chinese academy of wangzhe07@iscas ac cn ABSTRACT Nowadays people are inundated by choices. Personal- Collaborative filtering(CF) is a method for personal- ized recommendation is a solution to this problem. Var- ized recommendation. The sparsity of rating data se ious kinds of recommender systems are employed for riously impairs the quality of CF's recommendation. better user experience. Collaborative filtering 4, 12 is Meanwhile, there is more and more tag information gen- one of the best techniques of choice therein. This tech ated by online users that implies their preferences. nique tries to identify users that have relevant interests Exploiting these tag data is a promising means to al- by calculating similarities among user profiles. The idea leviate the sparsity problem. Although the intention is is that it may be of benefit to one's search for informa- raight-forward, there's no existed solution that makes tion to consult the behavior of other users who share full use of tags to improve the recommendation qual- the same or relevant interests ity of traditional rating-based collaborative filtering ap- proaches. In this paper, we propose a novel approach Because collaborative filtering recommendation depends to fuse a tag-based neighborhood method into the tradi- on the preference of the users with the same or rele- tional rating-based CF. Tag-based neighborhood method vant interests, the similarity computation imposes sig employed to find similar users and items. These nificant influence on the quality of recommendation neighborhood information helps the sequent Cf pro- Early item-based and user-based collaborative filtering cedure produce higher quality recommendations. The approaches find similar users or items(neighbors) by experiments show that our approach outperforms the calculating Pearson correlation coefficient (23. These state-of-the-art ones approaches are efficient and effective. But simply com- paring the rating records of different users or items can- ACM Classification Keyword not help to find the best neighbors. If a user has few H.3.3 Information Storage and Retrieval: Information ratings for items or this user only gives all his/her rat- Search and Retrieval-Information filtering ings to the unpopular ones, it will be difficult for those approaches to find the proper neighbors General terms gorithms, Experimentation Recently, matrix factorization approaches earn popularity because of their higher recommendation ity and smaller online costs. One of the most signif Author Keywords cant differences from early approaches is that they ex- Tags, Latent Dirichlet Allocation(LDA), Collaborative tract the "features"of the users and the items. By this Filtering, Neighborhood Method way, they decompose the original preference matrix into several low rank approximates or the items. ev- INTRODUCTION ery feature reflects the preference by a group of similar users. For the users, every feature reflects their pref- erence for a collection of similar items. By virtue of Permission to make d extracting users'and items'features, matrix factoriza or hard copies of all or part of this work for tion approaches are able to find bett ts an not made or distributed for profit or commercial advantage and that copies hence produce better recommendations bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific Despite the merits mentioned before, the existing ma- ermission and/or a fee Workshop SRS'10, February 7, 2010 Hong Kong, China trix factorization approaches 6, 7, 8, 16, 26 fail to ex Copyright2010ACM978-1-60558-995-4.s10.00Tags Meet Ratings: Improving Collaborative Filtering with Tag-Based Neighborhood Method Zhe Wang National Engineering Research Center of Fundamental Software, Institute of Software, Chinese Academy of Sciences Graduate University of Chinese Academy of Sciences wangzhe07@iscas.ac.cn Yongji Wang National Engineering Research Center of Fundamental Software, Institute of Software, Chinese Academy of Sciences ywang@itechs.iscas.ac.cn Hu Wu National Engineering Research Center of Fundamental Software, Institute of Software, Chinese Academy of Sciences wuhu@itechs.iscas.ac.cn ABSTRACT Collaborative filtering (CF) is a method for personal￾ized recommendation. The sparsity of rating data se￾riously impairs the quality of CF’s recommendation. Meanwhile, there is more and more tag information gen￾erated by online users that implies their preferences. Exploiting these tag data is a promising means to al￾leviate the sparsity problem. Although the intention is straight-forward, there’s no existed solution that makes full use of tags to improve the recommendation qual￾ity of traditional rating-based collaborative filtering ap￾proaches. In this paper, we propose a novel approach to fuse a tag-based neighborhood method into the tradi￾tional rating-based CF. Tag-based neighborhood method is employed to find similar users and items. These neighborhood information helps the sequent CF pro￾cedure produce higher quality recommendations. The experiments show that our approach outperforms the state-of-the-art ones. ACM Classification Keywords H.3.3 Information Storage and Retrieval: Information Search and Retrieval-Information filtering General Terms Algorithms, Experimentation Author Keywords Tags, Latent Dirichlet Allocation (LDA), Collaborative Filtering, Neighborhood Method INTRODUCTION Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, to republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Workshop SRS’10, February 7, 2010 Hong Kong, China Copyright 2010 ACM 978-1-60558-995-4... $10.00 Nowadays people are inundated by choices. Personal￾ized recommendation is a solution to this problem. Var￾ious kinds of recommender systems are employed for better user experience. Collaborative filtering [4, 12] is one of the best techniques of choice therein. This tech￾nique tries to identify users that have relevant interests by calculating similarities among user profiles. The idea is that it may be of benefit to one’s search for informa￾tion to consult the behavior of other users who share the same or relevant interests. Because collaborative filtering recommendation depends on the preference of the users with the same or rele￾vant interests, the similarity computation imposes sig￾nificant influence on the quality of recommendation. Early item-based and user-based collaborative filtering approaches find similar users or items (neighbors) by calculating Pearson correlation coefficient [23]. These approaches are efficient and effective. But simply com￾paring the rating records of different users or items can￾not help to find the best neighbors. If a user has few ratings for items or this user only gives all his/her rat￾ings to the unpopular ones, it will be difficult for those approaches to find the proper neighbors. Recently, matrix factorization approaches earn more popularity because of their higher recommendation qual￾ity and smaller online costs. One of the most signifi- cant differences from early approaches is that they ex￾tract the “features” of the users and the items. By this way, they decompose the original preference matrix into several low rank approximates [15]. For the items, ev￾ery feature reflects the preference by a group of similar users. For the users, every feature reflects their pref￾erence for a collection of similar items. By virtue of extracting users’ and items’ features, matrix factoriza￾tion approaches are able to find better neighbors and hence produce better recommendations. Despite the merits mentioned before, the existing ma￾trix factorization approaches [6, 7, 8, 16, 26] fail to ex-