SpeGCL: Self-Supervised Graph Spectrum Contrastive Learning Without Positive Samples.

Area of Science:

• Graph representation learning
• Machine learning on graphs
• Spectral graph theory

Background:

• Graph contrastive learning (GCL) effectively handles noisy graph data but struggles with fine-grained changes.
• Traditional graph convolutional networks (GCNs) preserve smooth features, limiting their ability to capture subtle graph variations.

Purpose of the Study:

• To develop a novel self-supervised graph spectrum contrastive learning framework (SpeGCL).
• To address the limitations of GCNs in capturing fine-grained changes in graph-structured data.

Main Methods:

• Constructed a Fourier graph neural network (FourierGNN) using stacked Fourier graph operations (FGO) layers for spectral analysis.
• Proposed a contrastive strategy focusing on maximizing differences in high-frequency information between augmented graph views.
• Provided theoretical justification for the effectiveness of using only negative samples in the contrastive learning objective.

Main Results:

• SpeGCL effectively captures fine-grained changes by analyzing graph frequency components.
• The proposed contrastive strategy, emphasizing high-frequency differences, leads to performance gains.
• SpeGCL demonstrated superior performance over state-of-the-art GCL methods in unsupervised, transfer, and semi-supervised learning tasks.

Conclusions:

• SpeGCL offers a powerful new approach for graph contrastive learning by incorporating spectral analysis.
• The framework's ability to leverage frequency information enhances its robustness and performance on diverse graph learning tasks.