TP-GCL: graph contrastive learning from the tensor perspective.

Area of Science:

• Machine Learning
• Graph Theory
• Data Science

Background:

• Graph Neural Networks (GNNs) excel at graph data analysis but struggle with complex structures and sparse labels.
• Limitations in information capture and generalization hinder traditional GNNs in practical applications.

Purpose of the Study:

• To overcome limitations of traditional GNNs in modeling complex graph structures.
• To address challenges posed by sparse labeling in graph datasets.
• To enhance the generalization capabilities of GNNs.

Main Methods:

• Proposed TP-GCL, a novel graph contrastive learning method with a tensor perspective.
• Transformed graphs into hypergraphs via clique expansion.
• Utilized high-order adjacency tensors to represent hypergraphs and capture complex structural information.
• Implemented a contrastive learning framework comparing original graphs with tensorized hypergraphs.

Main Results:

• TP-GCL demonstrated significant performance improvements over baseline methods on multiple public datasets.
• The method showed enhanced generalization capabilities.
• Effectiveness in handling complex graph structures and sparse labeled data was confirmed.

Conclusions:

• TP-GCL effectively extracts crucial structural features from graph data.
• The tensor-based contrastive learning approach offers a powerful solution for advanced GNN applications.
• This method advances GNN performance, especially in scenarios with complex structures and limited labels.