Graph Pooling with Local Cluster Selection

Graph poolings in GNNs are a family of operations which take graphs as inputs and produce coarsened graphs as output. Modern graph poolings are trainable and closely related to GNNs, which learn to pool graphs under different assumptions. Though there are various assumptions, the procedure of generating pooled graphs is relatively similar and limited. This work formalizes a novel procedure of pooling graphs, along with a graph pooling approach for average situations.