Abstract
Graph Neural Networks (GNNs) have achieved unprecedented success inidentifying categorical labels of graphs. However, most existing graphclassification problems with GNNs follow the protocol of balanced datasplitting, which misaligns with many real-world scenarios in which some classeshave much fewer labels than others. Directly training GNNs under thisimbalanced scenario may lead to uninformative representations of graphs inminority classes, and compromise the overall classification performance, whichsignifies the importance of developing effective GNNs towards handlingimbalanced graph classification. Existing methods are either tailored fornon-graph structured data or designed specifically for imbalanced nodeclassification while few focus on imbalanced graph classification. To this end,we introduce a novel framework, Graph-of-Graph Neural Networks (G$^2$GNN),which alleviates the graph imbalance issue by deriving extra supervisionglobally from neighboring graphs and locally from stochastic augmentations ofgraphs. Globally, we construct a graph of graphs (GoG) based on kernelsimilarity and perform GoG propagation to aggregate neighboring graphrepresentations. Locally, we employ topological augmentation via masking nodefeatures or dropping edges with self-consistency regularization to generatestochastic augmentations of each graph that improve the model generalibility.Extensive graph classification experiments conducted on seven benchmarkdatasets demonstrate our proposed G$^2$GNN outperforms numerous baselines byroughly 5\% in both F1-macro and F1-micro scores. The implementation ofG$^2$GNN is available athttps://github.com/YuWVandy/G2GNN}{https://github.com/YuWVandy/G2GNN