Modeling Implicit Feedback and Latent Visual Features for Machine-Learning Based Recommendation

摘要

To leverage the rapid accumulation of rich media on the Internet, this paper proposes a Multi-View Bayesian Personalized Ranking (MVBPR) recommendation model, which combines visual and textual content, along with uncertainty modeling of consumer preference in form of implicit feedback and visual representation in form of latent factors. MVBPR is a machine-leaning framework integral of deep-learning (i.e., SCAE) and topic modeling (i.e., LDA) strategies to fuse image and text information. Moreover, extensive experiments demonstrate MVBPR's advantages over baseline models, including its superiority in dealing with the cold start situation.