Negative-Free Self-Supervised Gaussian Embedding of Graphs.

Area of Science:

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

Background:

• Graph Contrastive Learning (GCL) learns node representations without labels using alignment and uniformity objectives.
• Existing GCL methods rely heavily on negative samples for uniformity, leading to high computational and memory demands.
• Negative sampling in GCL can cause representation collapse and class collision issues.

Purpose of the Study:

• To propose a novel negative-free objective for Graph Contrastive Learning to achieve representation uniformity.
• To eliminate the reliance on negative samples, thereby reducing computational and memory overhead.
• To maintain or improve the performance of GCL methods while addressing their inherent limitations.

Main Methods:

• Introduced a negative-free objective inspired by the uniform distribution of points from a normalized isotropic Gaussian.
• Minimized the distance between learned representations and an isotropic Gaussian distribution to enforce uniformity.
• Eliminated the need for parameterized mutual information estimators, additional projectors, and asymmetric network structures.

Main Results:

• Achieved competitive performance on seven graph benchmarks compared to existing GCL methods.
• Demonstrated significant reductions in computational demands, memory consumption, and training time.
• Successfully promoted uniformity in node representations without utilizing negative samples.

Conclusions:

• The proposed negative-free objective effectively achieves representation uniformity in Graph Contrastive Learning.
• This approach offers a more efficient and less resource-intensive alternative to traditional GCL methods.
• The findings suggest a new direction for developing scalable and effective graph self-supervised learning frameworks.