Select Your Own Counterparts: Self-Supervised Graph Contrastive Learning With Positive Sampling.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Self-Supervised Learning

Background:

• Contrastive learning (CL) is a powerful self-supervised learning technique.
• Sampling bias is a significant limitation in CL, hindering performance.
• Existing debiasing methods like hard negative mining (HNM) and supervised CL (SCL) are not fully effective for graph CL (GCL).

Purpose of the Study:

• To introduce a novel learning paradigm, graph positive sampling (GPS), to address sampling bias in GCL.
• To develop new contrastive objectives that enhance positive sample fusion and representative selection in the semantic space.
• To improve the performance and applicability of GCL models.

Main Methods:

• Proposing graph positive sampling (GPS), a new paradigm for GCL.
• Utilizing four complementary similarity measurements (node centrality, topological distance, neighborhood overlapping, semantic distance) to select positive counterparts for nodes.
• Developing three contrastive objectives to fuse positive samples and improve semantic representation.
• Implementing and evaluating three node-level GCL models incorporating GPS.

Main Results:

• GPS demonstrates superior performance compared to state-of-the-art (SOTA) baselines and existing debiasing methods in GCL.
• Extensive experiments on public datasets validate the effectiveness of GPS.
• GPS proves to be versatile, adaptive, and flexible for GCL applications.

Conclusions:

• Graph positive sampling (GPS) effectively mitigates sampling bias in graph contrastive learning.
• GPS offers a label-free approach, enabling preprocessing applications and enhancing GCL model performance.
• The proposed method represents a significant advancement in self-supervised learning for graph data.