Effective metric learning with co-occurrence embedding for collaborative recommendations.

In recommender systems, matrix factorization and its variants have grown up to be dominant in collaborative filtering due to their simplicity and effectiveness. In matrix factorization based methods, dot product which is actually used as a measure of distance from users to items, does not satisfy the inequality property, and thus may fail to capture the inner grained preference information and further limits the performance of recommendations. Metric learning produces distance functions that capture the essential relationships among rating data and has been successfully explored in collaborative recommendations. However, without the global statistical information of user-user pairs and item-item pairs, it makes the model easy to achieve a suboptimal metric. For this, we present a co-occurrence embedding regularized metric learning model (CRML) for collaborative recommendations. We consider the optimization problem as a multi-task learning problem which includes optimizing a primary task of metric learning and two auxiliary tasks of representation learning. In particular, we develop an effective approach for learning the embedding representations of both users and items, and then exploit the strategy of soft parameter sharing to optimize the model parameters. Empirical experiments on four datasets demonstrate that the CRML model can enhance the naive metric learning model and significantly outperforms the state-of-the-art methods in terms of accuracy of collaborative recommendations.