Hierarchically Contrastive Hard Sample Mining for Graph Self-Supervised Pretraining.

Area of Science:

• Graph representation learning
• Self-supervised learning
• Machine learning

Background:

• Contrastive learning is a key technique for graph self-supervised pretraining (GSP).
• Existing GSP methods struggle with sample imbalance and limited contrast patterns, hindering representation quality.
• Maximizing mutual information (MI) in GSP can lead to overlooking crucial 'hard' samples.

Purpose of the Study:

• To address limitations in current contrastive GSP algorithms.
• To propose a novel GSP algorithm, Hierarchically Contrastive Hard Sample Mining (HCHSM).
• To enhance graph representation learning by focusing on hard samples and integrating multi-level graph features.

Main Methods:

• Developed HCHSM, a novel GSP algorithm.
• Implemented a MI-based hard sample selection (MHSS) module for hierarchical filtering.
• Introduced a hierarchically contrastive scheme for multi-level feature integration.

Main Results:

• HCHSM outperforms existing methods on node classification and clustering tasks across seven benchmark datasets.
• The MHSS module effectively filters easy nodes, focusing on harder samples.
• Hierarchical contrastive learning enhances discrimination of hard samples and improves graph embedding quality.

Conclusions:

• HCHSM offers a more effective approach to graph self-supervised pretraining.
• The proposed method successfully addresses sample imbalance and limited contrast patterns in GSP.
• HCHSM demonstrates superior performance in downstream graph analysis tasks.