Research On Recommendation Algorithm With Collaborative Social Topic Regression

In the context of current recommendation systems, there are relationships betweenmost kinds of items. They are either obvious or implicit, for example, the friendship orfollow relationship between users on social network and the citations between academicpapers are obvious, relationship between movies is implicit. Additively, these itemsalso have rich text semantic information, such as the title and abstract of papers, tweetsand re-tweets from users on micro-blogs, description texts of movies. This workproposed to utilize the relationships and sematic information of items to boost therecommendation performance.In this work, we surveyed the state-of-art models and algorithms ofrecommendation. We proposed collaborative social topic regression model, which fusestopic model regularizing matrix factorization for both user action matrix and itemrelationship matrix. And we justify collaborative social topic regression on typicalrecommendation situation and cold-start recommendation situation in the context ofcelebrity recommendation on Twitter and TencentWeibo. We also show therecommendation difference for different users. Finally we explore and verify theeffectiveness of our model on mining user interests and gaining recommendationinterpretability.The contribution of this paper are listed as follows:(1) We proposed CollaborativeSocial Topic Regression (CSTR) to model the user behaviors, relationship network andsemantic information. Specifically, we combine topic model with matrix factorizationfor relational network matrix of items and user action matrix to from a unifiedhierarchical Bayesian model, which can address the warm-start and cold-start problemssimultaneously. And wealso proposed a modeling strategy, which incorporates differentconfidences for different dyadic contexts.(2) For proposed model CSTR, wedeveloped a fast coordinate ascent algorithm to optimize the corresponding model. Thealgorithm can gain the local optimized solution. The time complexity is only linearwith the number of users and items.(3) We conducted experiments on two real-worlddatasets from Twitter and TencentWeibo. The experiment result show that CSTRachieves a higher performance and provide more effective results than the state-of-art methods especially when recommending new items. In the context of cold-startsituation, CSTR has an improvement of72%and60%for Recall,136%and121%forAverage Precision on Twitter and Tencent Weibo respectively.(4) We showed thatCSTR captures user interests more precisely and gives better recommendationinterpretability via Twitter Dataset.