GDPool Adaptive Granularity for Hierarchical Graph Neural Networks

GDPool replaces fixed pooling ratios in hierarchical GNNs with a self-attention mechanism that selects graph-specific anchors. This approach mitigates gradient conflicts between layers and adapts pooling granularity to the unique structure of each individual graph.