Collaborative Generative Adversarial Network for Recommendation Systems

Recommendation systems have been a core part of daily Internet life. Conventional recommendation models hardly defend adversaries due to the natural noise like misclicking. Recent researches on GAN-based recommendation systems can improve the robustness of the learning models, yielding the state-of-the-art performance. The basic idea is to adopt an interplay minimax game on two recommendation systems by picking negative samples as fake items and employ reinforcement learning policy. However, such strategy may lead to mode collapse and result in high vulnerability to adversarial perturbations on its model parameters. In this paper, we propose a new collaborative framework, namely Collaborative Generative Adversarial Network (CGAN), which adopts Variational Auto-encoder (VAE) as the generator and performs adversarial training in the continuous embedding space. The formulation of CGAN has two advantages: 1) its auto-encoder takes the role of generator to mimic the true distribution of users preferences over items by capturing subtle latent factors underlying user-item interactions; 2) the adversarial training in continuous space enhances models robustness and performance. Extensive experiments conducted on two real-world benchmark recommendation datasets demonstrate the superior performance of our CGAN in comparison with the state-of-the-art GAN-based methods.