Adversarial Training of Deep Autoencoders Towards Recommendation Tasks

The increasing popularity of applying deep neural networks in collaborative filtering (CF) indicates its achievements in item recommendation tasks. More specifically, this paper focuses on Generative Adversarial Networks (GAN), where a generator and a discriminator play a minimax game so that the generator eventually learns to generate realistic data. While there are some existing GAN-based CF methods, we point out an issue regarding the employment of standard matrix factorization as their basic model, which is linear and difficult to capture the non-linear, subtle latent factors underlying user-item interactions. We believe that the extension to deep neural networks, which have demonstrated their high potential to figure out non-linear latent factors, would bring more impressive recommendation results. In this paper, we propose a novel GAN-based CF method, where Autoencoder, one of the most successful deep neural networks, takes a role of a generator to capture the true distribution of users' preferences over items. We also employ Bayesian personalized ranking (BPR) as our discriminative model so as to further improve the accuracy of our method. Experimental results on three real-life datasets demonstrate the superiority of our proposed method over state-of-the-art methods.