Basis Learning Autoencoders for Hybrid Collaborative Filtering in Cold Start Setting

Abstract

In recent years, most recommender systems rely on collaborative filtering (CF) based on matrix factorization (MF) that can predict unknown ratings by completing a rating matrix. However, this approach cannot be used for the cold start where no rating information is available for a given user or item. To address this problem, we develop a new hybrid CF (HCF) technique incorporating CF with content information. The proposed HCF is based on an auto-encoder (AE) consisting of a nonlinear encoder and a linear decoder. This type of AE is called the basis learning AE (BAE), because it can learn the basis of the row space of a sparse input matrix by its encoder. In the proposed scheme, the input to the BAE is a content augmented rating matrix; the BAE learns the basis of the row space of a given rating matrix, which is a subset of the basis of the content augmented rating matrix, and recovers each row of the rating matrix by a linear combination of the learned basis. Unlike most existing HCF schemes, our model does not incorporate additional content-based objective terms; yet extensive experiments on real-world datasets show that the proposed HCF can significantly advance the state-of-the-art.