Unifying Graph Contrastive Learning via Graph Message Augmentation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Graph contrastive learning typically relies on Graph Data Augmentation (GDA) before training Graph Neural Networks (GNNs).
• Existing GDAs involve modifications like dropping or perturbing graph components, but a universal and effective approach is lacking.
• The effectiveness of GDA is crucial for the performance of graph contrastive learning methods.

Purpose of the Study:

• To address the lack of a universal and effective Graph Data Augmentation (GDA) technique for diverse graph data.
• To introduce a novel Graph Message Augmentation (GMA) scheme that unifies and improves existing GDA methods.
• To propose a unified graph contrastive learning framework, Graph Message Contrastive Learning (GMCL), leveraging the proposed GMA.

Main Methods:

• Introduction of a graph message representation for graph data.
• Development of Graph Message Augmentation (GMA), a universal scheme reformulating existing GDAs.
• Implementation of attribution-guided universal GMA within a unified graph contrastive learning framework (GMCL).
• Facilitation of mixup augmentation for graph data, which is typically challenging.

Main Results:

• GMA provides a new perspective for understanding existing GDAs.
• GMA offers a universal and more effective graph data augmentation strategy for self-supervised learning.
• The proposed GMCL framework demonstrates effectiveness across various graph learning tasks.
• Experiments validate the benefits of both GMA and GMCL.

Conclusions:

• Graph Message Augmentation (GMA) presents a significant advancement in graph data augmentation.
• The unified Graph Message Contrastive Learning (GMCL) framework effectively leverages GMA for improved graph representation learning.
• The proposed methods offer a more universal and efficient approach to graph self-supervised learning.