Graph partitioning and graph neural network based hierarchical graph matching for graph similarity computation

Graph similarity computation aims to predict a score between one pair of graphs facilitate downstream applications, such as finding the most similar chemical compounds query compound or Fewshot 3D Action Recognition. Recently, some graph models based on neural networks have been proposed, which are either graph-level interaction node-level comparison. However, when number nodes in increases, it will inevitably bring about reduced representation ability high cost. Motivated by this observation, we propose partitioning and network-based model, called PSimGNN, effectively resolve issue. Specifically, each input is partitioned into set subgraphs extract local structural features directly. Next, novel network with an attention mechanism designed map subgraph embedding vector. Some these pairs automatically selected for comparison supplement subgraph-level fine-grained information. Finally, coarse-grained information among different integrated final score. Experimental results datasets sizes demonstrate that PSimGNN outperforms state-of-the-art methods tasks using approximate Edit Distance (GED) metric.