We propose an end-to-end learning-based approximate method for subgraph matching task, called subgraph matching network (Sub-GMN). First, Sub-GMN uses graph representation learning to map nodes to node-level embedding, and then combines metric learning and attention mechanisms to model the relationship between matched nodes in the data graph and query graph. Compared with the previous GNNs-based method for subgraph matching task, Sub-GMN can obtain the node-to-node matching relationships and allow varying the input composed of query graph and data graph in the test phase, while previous GNNs-based methods for subgraph matching task can only match a fixed and unchanged subgraph and cannot output the node-to-node matching relationships. In this paper, there are two contribution. The first contribution is that Sub-GMN is the first learning based methods for subgraph matching task and can output node-to-node matching relationships. To our best knowledge, no learning based methods have been proposed in formal journals that match subgraphs, and output node-to-node matching relationships, while allow varying query and data graphes for subgraph matching task. The second contribution is that Sub-GMN has achieved better experimental results than previous GNNs-based method for subgraph matching task from the perspective of accuracy and running time.

中文翻译：

---

Sub-GMN：子图匹配网络模型

我们为子图匹配任务提出了一种基于端到端学习的近似方法，称为子图匹配网络（Sub-GMN）。首先，Sub-GMN使用图表示学习将节点映射到节点级嵌入，然后结合度量学习和注意机制来对数据图和查询图中的匹配节点之间的关系进行建模。与以前的基于GNN的子图匹配方法相比，Sub-GMN可以获取节点到节点的匹配关系，并允许在测试阶段更改由查询图和数据图组成的输入，而以前的基于GNN的方法用于测试阶段。子图匹配任务只能匹配固定且不变的子图，不能输出节点到节点的匹配关系。本文有两个贡献。第一个贡献是Sub-GMN是第一个基于学习的子图匹配任务方法，可以输出节点到节点的匹配关系。据我们所知，在正式期刊中没有提出基于学习的方法来匹配子图，并输出节点到节点的匹配关系，同时允许为子图匹配任务使用不同的查询和数据图。第二个贡献是，从准确性和运行时间的角度来看，Sub-GMN取得了比以前的基于GNNs的子图匹配任务更好的实验结果。同时允许用于子图匹配任务的变化查询和数据图。第二个贡献是，从准确性和运行时间的角度来看，Sub-GMN取得了比以前的基于GNNs的子图匹配任务更好的实验结果。同时允许用于子图匹配任务的变化查询和数据图。第二个贡献是，从准确性和运行时间的角度来看，Sub-GMN取得了比以前的基于GNNs的子图匹配任务更好的实验结果。