Robust Recommendation with Adversarial Gaussian Data Augmentation

Recommender system holds the promise of accurately understanding and estimating the user preferences. However, due to the extremely sparse user-item interactions, the learned recommender models can be less robust and sensitive to the highly dynamic user preferences and easily changed recommendation environments. To alleviate this problem, in this paper, we propose a simple yet effective robust recommender framework by generating additional samples from the Gaussian distributions. In specific, we design two types of data augmentation strategies. For the first one, we directly produce the data based on the original samples, where we simulate the generation process in the latent space. For the second one, we firstly change the original samples towards the direction of maximizing the loss function, and then produce the data based on the altered samples to make more effective explorations. Based on both of the above strategies, we leverage adversarial training to optimize the recommender model with the generated data which can achieve the largest losses. In addition, we theoretically analyze our framework, and find that the above two data augmentation strategies equal to impose a gradient based regularization on the original recommender models. We conduct extensive experiments based on six real-world datasets to demonstrate the effectiveness of our framework.