Group Contrastive Self-Supervised Learning on Graphs.

Area of Science:

• Machine Learning
• Graph Representation Learning
• Self-Supervised Learning

Background:

• Contrastive learning methods for graphs typically optimize two-view representations.
• Existing methods often compute a single graph-level representation, limiting captured characteristics.
• There is a need for methods that capture more abundant and diverse graph properties.

Purpose of the Study:

• To propose a group contrastive learning framework for self-supervised graph representation learning.
• To enable graph encoders to capture more abundant graph characteristics by contrasting in multiple subspaces.
• To develop principled objectives for learning diverse and informative representations.

Main Methods:

• The proposed framework embeds graphs into multiple subspaces.
• Each representation is prompted to encode specific graph characteristics.
• Principled objectives capture intra-space and inter-space relations within groups.
• An attention-based group generator computes representations capturing different substructures.
• Two existing methods are extended to GroupCL and GroupIG within the framework.

Main Results:

• The group contrastive learning framework achieves a significant performance boost across various datasets.
• Experimental results demonstrate improved graph representation learning.
• Qualitative analysis confirms that generated features capture specific graph characteristics.

Conclusions:

• The group contrastive learning framework effectively enhances self-supervised graph representation learning.
• Contrasting graphs in multiple subspaces leads to richer and more informative representations.
• The proposed method offers a promising direction for future research in graph representation learning.