Recognizing Predictive Substructures with Subgraph Information Bottleneck

The emergence of Graph Convolutional Network (GCN) has greatly boosted the progress graph learning. However, two disturbing factors, noise and redundancy in data, lack interpretation for prediction results, impede further development GCN. One solution is to recognize a predictive yet compressed subgraph get rid obtain interpretable part graph. This setting similar information bottleneck (IB) principle, which less studied on graph-structured data Inspired by IB we propose novel (SIB) framework such subgraphs, named IB-subgraph. intractability mutual discrete nature makes objective SIB notoriously hard optimize. To this end, introduce bilevel optimization scheme coupled with estimator irregular graphs. Moreover, continuous relaxation selection connectivity loss stabilization. We theoretically prove error bound our estimation noise-invariant Extensive experiments learning large-scale point cloud tasks demonstrate superior property