Local structure-aware graph contrastive representation learning.

Area of Science:

• Graph representation learning
• Machine learning
• Artificial intelligence

Background:

• Traditional Graph Neural Networks (GNNs) are limited by label information.
• Graph Contrastive Learning (GCL) methods address label issues but often focus on global or first-order neighborhood structures.
• A gap exists in effectively modeling local, multi-view structural information for GCL.

Purpose of the Study:

• To propose a novel Local Structure-aware Graph Contrastive representation Learning (LS-GCL) method.
• To effectively model node structural information from multiple perspectives, going beyond first-order neighborhoods.
• To improve graph representation learning for downstream tasks like node classification and link prediction.

Main Methods:

• Construct semantic subgraphs beyond first-order neighbors for local view embeddings.
• Utilize a shared GNN encoder for both subgraph-level and global graph-level node embeddings.
• Employ a multi-level contrastive loss function to maximize common information across different views.
• Apply pooling functions to generate subgraph-level graph embeddings.

Main Results:

• The proposed LS-GCL method outperforms state-of-the-art graph representation learning approaches.
• Demonstrated superior performance on six benchmark datasets.
• Achieved significant improvements in both node classification and link prediction tasks.

Conclusions:

• LS-GCL effectively captures local structural information through multi-view contrastive learning.
• The method provides a robust framework for enhancing graph representation learning.
• LS-GCL offers a promising advancement for tasks involving complex graph data.