An Adversarial Deep Hybrid Model for Text-Aware Recommendation with Convolutional Neural Networks

Abstract

The standard matrix factorization methods for recommender systems suffer from data sparsity and cold-start problems. Thus, in real-world scenarios where items are commonly associated with textual data such as reviews, it becomes necessary to build a hybrid recommendation model that can fully utilize the text features. However, existing methods in this area either cannot extract good features from the texts due to their order–insensitive document modeling approaches or fail to learn the hybrid model in an effective way due to their complexity of inferring the latent vectors. To this end, we propose a deep hybrid recommendation model which seamlessly integrates matrix factorization with a Convolutional Neural Network (CNN), a powerful text feature extraction tool with the capability of detecting the information of word orders. Unlike previous works which use content features as prior knowledge to regularize the latent vectors, we combine CNN into MF in an additive manner to allow training CNN with direct learning signals. Furthermore, we propose an adversarial training framework to learn the hybrid recommendation model, where a generator model is built to learn the distribution over the pairwise ranking pairs while training a discriminator to distinguish generated (fake) and real item pairs. We conduct extensive experiments on three real-world datasets to demonstrate the effectiveness of our proposed model against state-of-the-art methods in various recommendation settings.