Latent space regularization for recommender systems

• Published in: Information sciences 2016-09, Vol.360, p.202-216
• Main Authors: Hong, Fu-Xing, Zheng, Xiao-Lin, Chen, Chao-Chao
• Format: Article
• Language: English
• Subjects: Diversity of recommendation; Factorization; Filtering; Filtration; Latent space regularization; Mathematical models; Ratings; Recommender systems; Regularization; Similarity; Tanimoto similarity
• ISSN: 0020-0255; 1872-6291
• DOI: 10.1016/j.ins.2016.04.042

The primary latent factor model cannot effectively optimize the user-item latent spaces because of the sparsity and imbalance of the rating data. Although existing studies have focused on exploring auxiliary information for users or items, few researchers have considered users and items jointly. For instance, social information is incorporated into models without considering the item side. In this paper, we introduce latent space regularization (LSR) and provide a general method to improve recommender systems (RSs) by incorporating LSR. We take the assumption that users prefer items that cover one or several topics that they are interested in, instead of all the topics, which reflects real-world situations. For instance, a user may focus on the humorous part of an item when he or she is at leisure time, regardless of the relevance of the item to his research topics. LSR operates from this assumption to account for both the user and item sides simultaneously. From another point of view, LSR is likely to improve the Tanimoto similarity of observed user-item pairs. As a result, LSR utilizes the number of ratings in a manner similar to weighted matrix factorization. We incorporate LSR into both the traditional collaborative filtering models that use only rating information and the collaborative filtering model that uses auxiliary content information as two examples. Experimental results from on two real-world datasets show not only the superiority of our model over other regularization models, but also its effectiveness and the possibility of incorporating it into various existing latent factor models.