[Paper Review] CGMN: A Contrastive Graph Matching Network for Self-Supervised Graph Similarity Learning

Di Jin, Luzhi Wang+5 more|arXiv (Cornell University)|May 30, 2022

Advanced Graph Neural Networks4 citations

TL;DR

This paper proposes CGMN, a contrastive graph matching network for self-supervised graph similarity learning that leverages cross-view and cross-graph interactions to improve node representation learning. By generating augmented views and jointly optimizing node-level and graph-level representations, CGMN achieves state-of-the-art performance on graph similarity tasks, outperforming existing unsupervised and supervised methods on multiple benchmarks including Aids700nef and BSD datasets.

ABSTRACT

Graph similarity learning refers to calculating the similarity score between two graphs, which is required in many realistic applications, such as visual tracking, graph classification, and collaborative filtering. As most of the existing graph neural networks yield effective graph representations of a single graph, little effort has been made for jointly learning two graph representations and calculating their similarity score. In addition, existing unsupervised graph similarity learning methods are mainly clustering-based, which ignores the valuable information embodied in graph pairs. To this end, we propose a contrastive graph matching network (CGMN) for self-supervised graph similarity learning in order to calculate the similarity between any two input graph objects. Specifically, we generate two augmented views for each graph in a pair respectively. Then, we employ two strategies, namely cross-view interaction and cross-graph interaction, for effective node representation learning. The former is resorted to strengthen the consistency of node representations in two views. The latter is utilized to identify node differences between different graphs. Finally, we transform node representations into graph-level representations via pooling operations for graph similarity computation. We have evaluated CGMN on eight real-world datasets, and the experiment results show that the proposed new approach is superior to the state-of-the-art methods in graph similarity learning downstream tasks.

Motivation & Objective

• To address the lack of self-supervised methods for joint graph representation learning and similarity scoring.
• To overcome limitations of clustering-based unsupervised methods that ignore cross-graph node matching and require manual cluster count selection.
• To design a contrastive learning framework that enhances node representation consistency across augmented views and captures inter-graph differences.
• To achieve high-performance graph similarity prediction without relying on expensive ground-truth labels.

Proposed method

• CGMN employs a siamese GNN architecture to process two graphs in a pair.
• It generates two augmented views for each graph to provide diverse contexts for contrastive learning.
• Cross-view interaction is used to enforce consistency in node representations across the two views of the same graph.
• Cross-graph interaction enables node-to-node matching between corresponding nodes in different graphs to capture structural and feature differences.
• Graph-level representations are derived via pooling operations on refined node embeddings for final similarity computation.
• The model is trained end-to-end using contrastive loss to maximize agreement between positive graph pairs and minimize it for negative pairs.

Experimental results

Research questions

• RQ1 Can a self-supervised framework effectively learn graph similarity without relying on labeled similarity scores?
• RQ2 How do cross-view and cross-graph interactions contribute to improved node and graph representation learning?
• RQ3 Does the proposed contrastive learning strategy outperform clustering-based unsupervised methods in graph similarity tasks?
• RQ4 How sensitive is the model's performance to hyperparameters such as feature masking and edge removal probabilities during augmentation?

Key findings

• On the Aids700nef dataset, CGMN achieves an MSE of 6.641×10⁻³, outperforming the next-best method by 24.13% in relative improvement.
• On the Linux1000 dataset, CGMN improves MSE by 12.28% compared to supervised models and 21.67% compared to unsupervised baselines.
• In the BSD task, CGMN achieves the highest AUC scores across all six datasets, with results consistently superior to existing unsupervised methods.
• Ablation studies confirm that removing either cross-view or cross-graph interaction leads to significant performance degradation, validating the necessity of both components.
• Parameter sensitivity analysis shows that excessive augmentation (e.g., >10% feature masking or edge removal) degrades performance due to structural distortion.
• The model generalizes well to unseen test graphs, demonstrating strong zero-shot transferability in graph similarity prediction.