Graph Classification Method Based on Wasserstein Distance

Abstract Graph classification is a challenging problem, which attracts more and more attention. The key to solving this problem is based on what metric to compare graphs, that is, how to define graph similarity. Common graph classification methods include graph kernel, graph editing distance, graph embedding and so on. We introduce a new graph similarity metric, namely GRD (Geometric gRaph Distance). Our model GRD is composed of three sub-modules, which capture the differences between the graph structures from different aspects. Finally, the graph distances defined by the three modules are fused to define the similarity between graphs. Experiments show that GRD is superior to the baseline methods on the benckmark datasets.