Neural Collaborative Filtering: Hybrid Recommendation Algorithm with Content Information and Implicit Feedback

Abstract

Collaborative filtering methods are widely used for recommender systems. However the performance degrades significantly because of highly sparse rating information in practical situations. Recently many hybrid models incorporating content information have been proposed to alleviate data sparsity. This paper first proposes the collaborative variational ranking model (CVRank) that combines variational autoencoder with pairwise ranking based collaborative filtering. The model learns latent factors of users and items from both rating and content information, and makes recommendation by calculating dot products between users and items. CVRank is suitable for other deep learning models, and can be easily extended to other multimedia other than text. Then this paper sums up similar models as neural collaborative filtering and presents a generic optimization criterion for them. Experiments show that CVRank achieves robust performance under different sparsity level, and neural collaborative filtering methods can gain greater recommendation accuracy improvement compared with pure collaborative filtering when the training data is sparser.