Weight-Aware Graph Contrastive Learning

In contrastive learning, samples usually have different contributions to optimization. This inference applies to the specific downstream tasks of applying contrastive learning to graph learning. Nevertheless, the nodes, i.e., samples, are equally treated in the conventional graph contrastive learning approaches. To address such a problem and better model discriminative information from graphs, we propose a novel graph contrastive learning approach called Weight-Aware Graph Contrastive Learning (WA-GCL). WA-GCL first pairs up the node feature as positive and negative pairs. The weight factors of these pairs are then determined based on their similarities. WA-GCL utilizes a specific weight-aware loss to effectively and efficiently learn discriminative representations from the pairs with their corresponding weight factors. Empirically, the experiments across multiple datasets demonstrate that WA-GCL outperforms the state-of-the-art methods in graph contrastive learning tasks. Further experimental studies verified the effectiveness of different parts of WA-GCL.