When Does Self-Supervision Help Graph Convolutional Networks?

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Neural Networks

Background:

• Self-supervision enhances representation learning in Convolutional Neural Networks (CNNs).
• Its application to Graph Convolutional Networks (GCNs) remains underexplored.
• This study addresses the gap by systematically investigating self-supervision for GCNs.

Purpose of the Study:

• To systematically explore and assess the incorporation of self-supervision into GCNs.
• To propose and evaluate novel self-supervised learning tasks for GCNs.
• To investigate the integration of multi-task self-supervision with graph adversarial training.

Main Methods:

• Elaborated three mechanisms for incorporating self-supervision into GCNs.
• Analyzed limitations of pretraining & finetuning and self-training.
• Proposed and compared three novel self-supervised learning tasks for GCNs.
• Integrated multi-task self-supervision into graph adversarial training.

Main Results:

• Self-supervision, when properly incorporated, improves GCN generalizability.
• Self-supervision enhances the robustness of GCNs.
• Novel self-supervised tasks showed promising results in theoretical and numerical comparisons.

Conclusions:

• Self-supervision is a beneficial technique for improving GCN performance.
• Careful design of self-supervised tasks and incorporation mechanisms is crucial.
• This work provides a foundation for future research in self-supervised GCNs.