Local Set Pooling (LSPool) is a hierarchical graph pooling that learns to coarsen input graphs from local set. LSPool is localized, adaptive and trainable. Its density depends on the density of input graphs.
LSPool assumes that for each node, the localized sub-graph contains the node and its neighbours determines whether this node presents in the pooled graph. This sub-graph is call local set. A local set is formed by a central node, its one-hop neighbours and edges connecting them.
LSPool uses two message passing layers(MP) to learn and construct the pooled graph. Local set score MP(LSSMP) learns a score for each local set. The nodes whose local sets have the top-k scores are kept. The node features are updated by local set collapse MP(LSCMP), which collapses the nodes of a local set into one node. If two nodes have intersected local set in original graph, there is an edge between them in the pooled graph.
LSSMP learns the score from three features: the central node attributes, the difference between central node and neighbours, the possible edge attributes.
LSSMP
where
LSCMP collapse a local set into a node in pooled graph by aggregating the node attributes with scores learned.
LSCMP