LDA For One-Class Collaborative Filtering In Recommender System

The explosive growth of the Internet has produced a large number of implicit feedbacks,such as the data generated by clicking,browsing and “favorite”.It is increasingly important to use these implicit feedbacks to infer user's potential preferences.Recently,LDA for one-class collaborative filtering has attracted more and more attention from both the industry and research communities.By introducing topic models into the recommendation process,we can guide users to the topics they may be interested in.Most existing recommendation models based on LDA have two limitations:(i)the previous works mainly focus on the user-based approach to learn the topic structure,while few works combine user-based method with item-based method or construct different preference context to improve recommendation performance;(ii)most recent works only exploit implicit feedbacks,but ignore the review of users to items.We notice that the review information can not only indicate users' preference intuitively,but also reflect certain attributes of items.To address the two limitations mentioned above,we propose three novel methods based on LDA for one-class collaborative filtering.Firstly,we design a novel hybrid LDA-based recommendation algorithm,i.e.,hybrid LDA for one-class collaborative filtering(Hybrid-LDA),which fuses the superiority of the user-based and item-based algorithms to enhance the recommendation performance.Secondly,we develop a novel algorithm called multi-segment preference context modeling based on LDA(MultiLDA),which aims to construct different segments of preference context based on users' ratings.Thirdly,we propose a novel method called Hybrid-CALDA,which is a contentaware LDA model that combines implicit feedbacks with review text data.Extensive empirical studies show that our Hybrid-LDA,Multi-LDA and Hybrid-CALDA can achieve better performance than the state-of-the-art methods in most cases.