Graph contrastive learning (GCL), which learns node representations by contrasting two augmented graphs, has attracted significant research interest. The core of GCL is to obtain high-quality positive and negative samples as contrastive pairs, thereby learning the latent semantic information of the original graph. Existing GCL methods typically use random augmentation strategies to generate augmented views. In existing methods, the embeddings of the same node across two augmented graph views are treated as positive pairs, while embeddings of different nodes are regarded as negative pairs. Nevertheless, random graph augmentations may inadvertently distort the inherent structural and semantic properties of the graph, thereby inducing semantic drift for the same node between the two augmented views. Additionally, two different nodes of the same class may have similar semantics after augmentation. Therefore, we argue that the previous method of defining positive and negative pairs faces a bottleneck due to potential inauthenticity between the pairs. To address this issue, we propose a method called Semantic-Aware Graph Contrastive Learning with Positive Sampling (GCLPS), which mines latent semantic information in the graph, incorporating truly semantically similar node pairs as positive pairs. This approach reduces the impact of false positive pairs and helps the GNN learn the semantic relationships between nodes. Specifically, in the topology aspect, we select one-hop neighbors of each node in the original graph as positive pairs. In the feature aspect, we represent nodes on augmented views using subgraph-level graph kernels and choose positive pairs based on their similarity. Finally, we combine these two aspects to form the final set of positive pairs for contrastive loss calculation. Extensive experiments on six benchmark datasets demonstrate that GCLPS achieves superior performance in various downstream tasks compared to state-of-the-art methods.


扫码关注我们
求助内容:
应助结果提醒方式:

