Collaborative filtering and content based hybrid models for recommending scientific articles

Abstract

Recommendation systems (RS) are programs that assist users in accessing infor mation in vast amount of data collections. In this thesis, we investigate hybrid models that use both implicit ratings such as tags, bookmarks or impressions, and content in formation such as user’s profile or item properties. In the literature, recommendation approaches that use such information are known as collaborative filtering (CF) and content-based methods, respectively. As computation methodology we investigate and compare two techniques, one is based on matrix decomposition and the other one is based on deep learning. As a matrix decomposition based approach, we investigate Bayesian nonnegative matrix factorization (BNMF), that we enhance using side infor mation, the titles and abstracts of scientific articles, besides the implicit rating matrix. As a deep learning method, we explore collaborative deep learning (CDL), which uses probabilistic matrix factorization as CF method and Bayesian stacked denoising au toencoder (SDAE) as content feature extraction. We apply these techniques in our experiments to a CiteULike dataset with a rating density of 0.22%. Our experimen tal results show that CDL is more effective than coupled BNMF on this dataset. In our opinion, CDL performs better due to its Bayesian SDAE component which has nonlinear and deep structure.