Generative and contrastive graph representation learning with message passing.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Artificial Intelligence

Background:

• Self-supervised graph representation learning (SSGRL) avoids manual labels for graph embeddings.
• Existing SSGRL methods are either generative (prone to poor quality) or contrastive (sensitive to augmentation and negative samples).
• Neither pure generative nor contrastive approaches adequately balance robustness and performance.

Purpose of the Study:

• To propose a novel SSGRL method, Contrastive Generative Message Passing Graph Learning (CGMP-GL).
• To integrate generative and contrastive learning paradigms for improved graph representation.
• To enhance node representation discriminability and model robustness.

Main Methods:

• CGMP-GL integrates contrastive learning into generative models and message aggregation.
• Employs cross-view multi-level contrast to fuse multi-granularity topology and feature information.
• Reconstructs masked node features and optimizes representations via self-supervised contrastive message passing.

Main Results:

• CGMP-GL demonstrates enhanced discriminability by aligning positive and separating negative samples.
• The method effectively integrates topological and feature information across different granularities.
• Extensive experiments confirm the effectiveness and robustness of CGMP-GL on multiple datasets and tasks.

Conclusions:

• CGMP-GL offers a robust and effective approach to self-supervised graph representation learning.
• The integrated generative and contrastive strategy overcomes limitations of existing methods.
• CGMP-GL significantly improves performance on various downstream graph tasks.