Hierarchical Glocal Attention Pooling for Graph Classification

Waqar Ali, Sebastiano Vascon, Thilo Stadelmann, Marcello Pelillo

Abstract: Highlights•Propose dual-fold pooling to capture global and local properties for classification.•Fold1 uses a developed rule-based method to identify overlapping nodes among cliques.•Develop dynamic scoring to rank the most informative global structures like cliques.•Fold2 uses LocalPool to refine cliques by focusing on key nodes within cliques.•The proposed method outperforms 11 state-of-the-art models on seven diverse datasets.