Collaborative Deep Embedding via Dual Networks

Despite the long history of research on recommender systems, current approaches still face a number of challenges in practice, e.g. the difficulties in handling new items, the high diversity of user interests, and the noisiness and sparsity of observations. Many of such difficulties stem from the lack of expressive power to capture the complex relations between items and users. This paper presents a new method to tackle this problem, called Collaborative Deep Embedding. In this method, a pair of dual networks, one for encoding items and the other for users, are jointly trained in a collaborative fashion. Particularly, both networks produce embeddings at multiple aligned levels, which, when combined together, can accurately predict the matching between items and users. Compared to existing methods, the proposed one not only provides greater expressive power to capture complex matching relations, but also generalizes better to unseen items or users. On multiple real-world datasets, this method outperforms the state of the art.