Deformable Convolutional Matrix Factorization for Document Context-Aware Recommendation in Social Networks

Abstract

The extreme sparsity of the rating data seriously affects the recommendation quality of the recommendation system. In order to alleviate the problem of data sparsity, some convolutional neural network (CNN)-based models make full use of text data to improve the recommendation accuracy. However, due to the inherent properties of the traditional convolutional network, it can only extract features in a fixed position, and rely on the primitive bounding box based feature extraction, thus ignoring the flexibility of the traditional convolution. In this paper, we adopt a flexible convolutional network called deformable convolutional network (DCN), which extends the convolution transformation model capability by adding an offset layer to the traditional convolution layer, and then propose a novel deformable convolutional network matrix factorization (DCNMF) recommendation model. Specifically, we combine the DCN with word embedding to deeply capture the contextual information of document and build a latent model, which is incorporated into the probabilistic matrix factorization (PMF) model to enhance the recommendation accuracy. We conduct extensive experiments on the real-world datasets, and the experimental results show that the DCNMF outperforms the compared benchmarks.