T3A: robust graph contrastive learning with tensorized augmentation, aggregation and approximation

Recently, graph contrastive learning (GraphCL) has received extensive attention owing to its unsupervised training paradigm and powerful representation learning capability. However, existing GraphCL methods have seldom been researched for robustness, and their performance degrades significantly against adversarial attacks. Therefore, it is significant to consider the robustness of GraphCL.In this paper, a robust GraphCL model with tensorized augmentation, aggregation, and approximation (T3A) is proposed to defend against attacks and noise.First, a tensorized multi-graph augmentation is designed to enable the encoder to derive more essential and intrinsic consistency across multiple augmented graphs, distinct from the single-graph augmentation used in existing studies, which is vulnerable to perturbations.Second, the augmented views are fed into a tensorized graph aggregator, which is trained using the extended contrastive loss in matrix form to fuse information from multiple graphs.Finally, tensorized low-rank approximation is employed for the tensor representation to further extract the low-rankness of the graph, thereby improving model robustness.Extensive experiments are conducted on five datasets under three different adversarial attacks to demonstrate the effectiveness and robustness of T3A.