[Paper Review] Sub-GMN: The Subgraph Matching Network Model

Zixun Lan, Limin Yu+4 more|arXiv (Cornell University)|Mar 31, 2021

Advanced Graph Neural Networks26 references5 citations

TL;DR

This paper proposes Sub-GMN, an end-to-end learning-based model for subgraph matching that uses graph representation learning, metric learning, and attention mechanisms to generate node-to-node matches between query and data graphs. It achieves 12.21% higher accuracy than GNN and 3.2% higher than FGNN on dataset 1, runs 20–40× faster than FGNN, and outputs a list of matched node pairs—unlike most prior GNN-based methods.

ABSTRACT

As one of the most fundamental tasks in graph theory, subgraph matching is a crucial task in many fields, ranging from information retrieval, computer vision, biology, chemistry and natural language processing. Yet subgraph matching problem remains to be an NP-complete problem. This study proposes an end-to-end learning-based approximate method for subgraph matching task, called subgraph matching network (Sub-GMN). The proposed Sub-GMN firstly uses graph representation learning to map nodes to node-level embedding. It then combines metric learning and attention mechanisms to model the relationship between matched nodes in the data graph and query graph. To test the performance of the proposed method, we applied our method on two databases. We used two existing methods, GNN and FGNN as baseline for comparison. Our experiment shows that, on dataset 1, on average the accuracy of Sub-GMN are 12.21\% and 3.2\% higher than that of GNN and FGNN respectively. On average running time Sub-GMN runs 20-40 times faster than FGNN. In addition, the average F1-score of Sub-GMN on all experiments with dataset 2 reached 0.95, which demonstrates that Sub-GMN outputs more correct node-to-node matches. Comparing with the previous GNNs-based methods for subgraph matching task, our proposed Sub-GMN allows varying query and data graphes in the test/application stage, while most previous GNNs-based methods can only find a matched subgraph in the data graph during the test/application for the same query graph used in the training stage. Another advantage of our proposed Sub-GMN is that it can output a list of node-to-node matches, while most existing end-to-end GNNs based methods cannot provide the matched node pairs.

Motivation & Objective

• To address the NP-complete subgraph matching problem in a scalable, learnable way.
• To enable end-to-end learning for subgraph matching with flexible query and data graph pairs during inference.
• To produce a list of node-to-node matches rather than just identifying a single subgraph.
• To improve accuracy and inference speed over existing GNN-based subgraph matching methods.
• To support varying query graphs at test time, unlike prior GNN methods that require fixed training queries.

Proposed method

• Uses graph representation learning to embed nodes in the data and query graphs into low-dimensional vectors.
• Applies metric learning to learn a similarity metric between node embeddings in the query and data graphs.
• Integrates an attention mechanism to dynamically weigh the importance of node pairs during matching.
• Combines node embeddings and learned similarity scores to predict node-to-node matches.
• Trains the model end-to-end using a differentiable loss function to optimize matching accuracy.
• Supports arbitrary query graphs at test time, enabling generalization beyond training-specific queries.

Experimental results

Research questions

• RQ1 Can a deep learning model achieve higher accuracy than existing GNN-based methods in subgraph matching?
• RQ2 Can the model generalize to unseen query graphs during inference, unlike prior GNN-based approaches?
• RQ3 Does the model produce a list of node-to-node matches rather than just a single subgraph?
• RQ4 Can the model maintain high speed while improving accuracy in subgraph matching?
• RQ5 How does the integration of metric learning and attention improve matching performance?

Key findings

• On dataset 1, Sub-GMN achieved 12.21% higher accuracy than GNN and 3.2% higher than FGNN on average.
• Sub-GMN ran 20–40 times faster than FGNN in average running time on dataset 1.
• On dataset 2, Sub-GMN achieved an average F1-score of 0.95, indicating high precision and recall in node matching.
• Sub-GMN successfully outputs a list of node-to-node matches, a capability absent in most existing end-to-end GNN-based methods.
• The model generalizes to varying query graphs at test time, unlike most prior GNN-based subgraph matching models that are fixed to training queries.
• The integration of metric learning and attention mechanisms significantly enhances matching performance and robustness.