Deep Neural Factorization Machine for Recommender System.

Factorization Machine (FM) and its many deep learning variants are widely used in Recommender Systems. Despite their success in many applications, there still remain inherent challenges. Most existing FM methods are incapable of capturing the similarity of features well and usually suffer from irrelevant features in terms of recommendation tasks. Hence, it is necessary to fully utilize the similarity interaction between different features. In this paper, we propose a Deep Neural Factorization Machine, named DNFM, which contains "wide" and "deep" parts based on Wide&Deep. In the wide part, we elaborately design a Dimension-weighted Factorization Machine (DwFM) to improve the original FM. DwFM assigns different weights vectors to feature interactions from different fields interactions. The "deep" part of DNFM is a neural network, which is used to capture the nonlinear information of the feature interactions. Then we unify both the wide and deep parts to comprehensive learn the feature interactions. Extensive experiments verify that DNFM can significantly outperform state-ofthe-art methods on two well-known datasets.