Exploring Attention and Self-Supervised Learning Mechanism for Graph Similarity Learning

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Science

Background:

• Graph similarity estimation is complex due to intricate graph structures.
• Existing frameworks struggle to unify cross-graph interactions, similarity matrix mapping, and self-supervised learning.

Purpose of the Study:

• To propose a unified self-supervised framework for graph similarity learning.
• To address limitations in learning cross-graph interactions and mapping similarity matrices.
• To establish an effective self-supervised learning mechanism for graph similarity.

Main Methods:

• Developed a unified self-supervised nodewise attention-guided graph similarity learning framework (SNA-GSL).
• Employed correlation-guided contrastive learning for node embeddings.
• Utilized multiple attention mechanisms within graph similarity learning for score prediction.

Main Results:

• SNA-GSL demonstrates superior performance on graph-graph regression and graph classification tasks.
• The framework effectively captures node embeddings and predicts similarity scores.
• Achieved state-of-the-art results, indicating strong generalization capabilities.

Conclusions:

• The proposed SNA-GSL framework offers a robust solution for graph similarity learning.
• The attention-guided and self-supervised mechanisms significantly enhance performance.
• The model's success in graph classification highlights its generalization ability.