Adversarial Graph Perturbations for Recommendations at Scale

Abstract

Graph Neural Networks (GNNs) provide a class of powerful architectures that are effective for graph-based collaborative filtering. Nevertheless, GNNs are known to be vulnerable to adversarial perturbations. Adversarial training is a simple yet effective way to improve the robustness of neural models. For example, many prior studies inject adversarial perturbations into either node features or hidden layers of GNNs. However, perturbing graph structures has been far less studied in recommendations. To bridge this gap, we propose AdvGraph to model adversarial graph perturbations during the training of GNNs. Our AdvGraph is mainly based on min-max robust optimization, where an universal graph perturbation is obtained through an inner maximization while the outer optimization aims to compute the model parameters of GNNs. However, direct optimizing the inner problem is challenging due to discrete nature of the graph perturbations. To address this issue, an unbiased gradient estimator is further proposed to compute the gradients of discrete variables. Extensive experiments demonstrate that our AdvGraph is able to enhance the generalization performance of GNN-based recommenders.