Multilevel Graph Matching Networks for Deep Graph Similarity Learning.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Computer Science

Background:

• Graph neural networks (GNNs) excel at node representation but struggle with graph similarity learning.
• Existing methods often focus on global or node-level interactions, neglecting crucial cross-level connections.
• There's a need for advanced models to capture complex relationships between graph structures.

Purpose of the Study:

• To propose a novel Multilevel Graph Matching Network (MGMN) for end-to-end graph similarity computation.
• To effectively learn cross-level interactions between nodes of one graph and the entire other graph.
• To address the lack of benchmark datasets by creating new datasets for graph-graph classification and regression.

Main Methods:

• Developed a framework integrating a node-graph matching network (NGMN) for cross-level interactions.
• Employed a siamese GNN to capture global-level interactions between two input graphs.
• Created and curated new datasets for evaluating graph similarity models.

Main Results:

• MGMN consistently outperformed state-of-the-art baseline models on graph-graph classification and regression tasks.
• The proposed model demonstrated superior robustness with increasing graph sizes compared to previous approaches.
• Experiments validated the effectiveness of capturing cross-level interactions for accurate graph similarity.

Conclusions:

• The Multilevel Graph Matching Network (MGMN) framework offers a significant advancement in graph similarity learning.
• MGMN's ability to integrate node-graph and graph-graph interactions leads to improved performance and robustness.
• The study provides valuable datasets and a robust model for future research in graph representation learning.