Improving One-Class Collaborative Filtering with Manifold Regularization by Data-driven Feature Representation

Yen-Chieh Lien, Pu-Jen Cheng

Abstract: When considering additional features of users or items in a recommendation system, previous work focuses mainly on manually incorporating these features into original models. In this paper, manifold regularization is introduced to the well-known one-class collaborative filtering problem. To fully benefit from large unlabeled data, we design a data-driven framework, which learns a representation function by not only transferring raw features of users or items into latent ones but also directly linking the relation between the latent features and user behaviors. The framework is expected to bring cluster hypothesis from machine learning to recommendation, that is, more similar transferred features can bring more similar user behaviors. The experiments have been conducted on two real datasets. The results demonstrate that the learned representation through our framework can boost prediction performance significantly.