Matrix Co-factorization for Recommendation with Rich Side Information and Implicit Feedback Yi Fang Department of Computer Science Purdue University West Lafayette, IN 47907, USA fangy@cs.purdue.edu Luo Si Department of Computer Science Purdue University West Lafayette, IN 47907, USA lsi@cs.purdue.edu ABSTRACT Most recommender systems focus on the areas of leisure ac- tivities. As the Web evolves into omnipresent utility, recom- mender systems penetrate more serious applications such as those in online scientific communities. In this paper, we investigate the task of recommendation in online scientific communities which exhibit two characteristics: 1) there ex- ists very rich information about users and items; 2) The users in the scientific communities tend not to give explicit ratings to the resources, even though they have clear pref- erence in their minds. To address the above two character- istics, we propose matrix factorization techniques to incor- porate rich user and item information into recommendation with implicit feedback. Specifically, the user information matrix is decomposed into a shared subspace with the im- plicit feedback matrix, and so does the item information ma- trix. In other words, the subspaces between multiple related matrices are jointly learned by sharing information between the matrices. The experiments on the testbed from an online scientific community (i.e., Nanohub) show that the proposed method can effectively improve the recommendation perfor- mance. Categories and Subject Descriptors H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval—Information Filtering General Terms Algorithms, Performance, Experimentation Keywords Matrix co-factorization, Implicit feedback, Rich side infor- mation 1. INTRODUCTION Recommender systems attempt to analyze user prefer- ences over items, and model the relationship between users 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. HetRec’11, October 27, 2011, Chicago, IL, USA Copyright 2011 ACM 978-1-4503-1027-7/11/10... $10.00. and items in order to generate meaningful recommendations to the users. Such systems have been ubiquitously adopted in many applications such as e-commerce, social bookmark- ing, and subscription based services. They provide person- alized recommendations which are especially important in markets where the variety of choices is large and the taste of the customer is important. Most recommender systems fo- cus on the areas of leisure activities such as art (e.g., movies and books), fashion (e.g., music and gaming), and food (e.g., restaurants). As the Web evolves into omnipresent utility, recommender systems penetrate more serious applications such as those in online scientific communities. In this paper, we investigate the task of recommendation in online scientific communities. In particular, our study is based on the Nanohub 1 website hosted by Purdue Univer- sity. Nanohub is an online scientific community for research, education and collaboration in nanotechnology. It comprises numerous resources with an active user base. These re- sources include lectures, seminars, tutorials, publications, events and so on. The task is to recommend relevant re- sources to the users. The scientific communities such as Nanohub exhibit two characteristics: 1) there exists very rich information about resources and users. Most resources contain detailed information such as titles, abstracts and tags. Many registered users also provide detailed profiles about themselves such as research interest, education and affiliation. This information is very indicative for recom- mendation and thus needs to be taken into consideration. 2) The users in the scientific communities tend not to give explicit ratings to the resources, even though they have clear preference in their minds. There only exists implicit user feedback such as the user clicks on resources. These two characteristics may also be noticeable in many other real- world applications, while they are more prominent in online scientific communities such as Nanohub. This paper proposes matrix co-factorization techniques to incorporate rich user and resource information into recom- mendation with implicit feedback. Specifically, the user in- formation matrix is decomposed into a shared subspace with the implicit feedback matrix, and so does the item informa- tion matrix. In other words, the subspaces between multiple related matrices are jointly learned by sharing information between the matrices. To reflect the confidence level on the implicit feedback, the binary elements in the implicit feed- back matrix are weighted according to the frequency of the feedback (for 1) or the user-resource content similarity (for 0). In sum, our main contribution is to factorize implicit 1 http://www.nanohub.org