Learning Continuous User and Item Representations for Neural Collaborative Filtering

Abstract

Collaborative filtering (CF) is one of the most successful approach commonly used by many recommender systems. Recently, deep neural networks (DNNs) have performed very well in collaborative filtering for recommendation. Conventional DNNs based CF methods directly model the interaction between user and item features by transforming users and items into binarized sparse vectors with one-hot encoding, then map them into a low dimensional space with randomly initialized embedding layers and automatically learn the representations in training process. We argue that randomly initialized embedding layers can not capture the contextual relations of user interactions. We propose an approach that uses the optimized representations of user and item generated by doc2vec algorithm to initialize embedding layers. Items with similar contexts (i.e., their surrounding click) are mapped to vectors that are nearby in the embedding space. Our experiments on three industry datasets show significant improvements, especially in high-sparsity recommendation scenarios.