Graph Neural Networks with Multi-granularity Pooling

Abstract

Graph Neural Networks (GNNs) are widely used in various tasks such as graph or node classification and achieved state-of-the-art results. However, current GNN models are typically using an inherently flat or single global pooling step to aggregate node features, which lack of semantic information. Here we propose MgPOOL that can generate multi-granularity representations of graphs and can be combine with multiple types of GNN models. MgPOOL can reduce the size of graph in an adaptive and learn a multi-granular cluster assignment for nodes at each layer, mapping the similar nodes into the same cluster, which then a coarse-grained input is constructed for the next layer. Here we combine several existing GNN models to demonstrate that multi-granularity node classification is possible. The experimental results are verified on several established graph classification benchmarks and achieving a new state-of-the-art on five common benchmark data sets. Furthermore, the method provides a better interpretability for deep GNN models.