Grouping Matrix Based Graph Pooling with Adaptive Number of Clusters

Graph pooling is a crucial operation for encoding hierarchical structures within graphs. Most existing graph approaches formulate the problem as node clustering task which effectively captures topology. Conventional methods ask users to specify an appropriate number of clusters hyperparameter, then assuming that all input graphs share same clusters. In inductive settings where could vary, however, model should be able represent this variation in its layers order learn suitable Thus we propose GMPool, novel differentiable architecture automatically determines based on data. The main intuition involves grouping matrix defined quadratic form operator, induces use binary classification probabilities pairwise combinations nodes. GMPool obtains operator by first computing matrix, decomposing it. Extensive evaluations molecular property prediction tasks demonstrate our method outperforms conventional methods.