From Graph Generation to Graph Classification

The graph classification task is to assign a discrete class label to an input graph. The dominant approach for neural graph classification is to compute an embedding for the input graph and perform the final classification in embedding space. The successful graph coarsening approach aggregates graph structural information at successively lower resolutions until a final embedding is obtained. Another direction for graph learning, so far unrelated, is graph generation. A graph generative model (GGM) aims to generate realistic graphs, often by sampling from a distribution over graphs. GGMs include the graph Variational Auto-Encoder (GVAE), auto-regressive methods, and most recently graph diffusion models.