Adversarial Collaborative Auto-encoder for Top-N Recommendation

Recently, deep learning-based recommendation models have been proved to have state-of-the-art recommendation accuracy. However, most of the existing work assume that user feedbacks are noise-free, on which the neural networks (NN) are trained. Although some methods apply man-made noises on the input data to train the networks more effectively (e.g. the collaborative denoising auto-encoder), the noises are randomly generated. To gain further improvements, we focus on boosting the overall recommendation performance through adversarial noises. We propose a general framework to adversarially train a NN-based item recommendation model. In particular, we select the collaborative auto-encoder model as an example and test our method on three public datasets. We show that our approach enhances both overall robustness and performance which outperforms competitive state-of-the-art item recommendation models.