CentralityPool: Centrality-aware Hierarchical Graph Pooling

Graph pooling is the backbone of graph classification that condenses graph information into a compact form. Current pooling methods often treat all node information irrespective of their importance, which limits their ability to capture either key graph substructures or node information. To this end, we propose CentralityPool, a hierarchical pooling method for sieving the most important nodes via centrality techniques, and conduct experiments over graph classification and node classification tasks. We employ four centrality techniques: Personalized PageRank, Katz, Total Communicability, and Eigenvector centrality to identify key nodes. Due to time constraints, we incorporate efficient approximation strategies to reduce computational complexity. Hence, at comparable computation time with SOTA methods, CentralityPool improves graph classification accuracy by up to 4% by average across eight benchmark datasets and achieves competitive node classification results across large-scale benchmark datasets.