Foundations for Robust yet Simple Sparse Hierarchical Pooling: A New Perspective on Sparse Graph Pooling

This work investigates Standard Sparse Pooling (SSP) methods within Graph Neural Networks, focusing on their effectiveness in preserving graph-level information while performing local pooling. We analyze the role of Selection and Reduction functions in SSP and introduce a new perspective that addresses the shortcomings of existing methods. We reveal that while SSP is simple, it has limitations in forming hierarchical representations, leading to potential over-representation in certain regions. This study provides foundational insights into achieving robust yet simple sparse pooling without unnecessary complexities.