Improving effectiveness and computational complexity of graph pooling

Investigate and develop graph pooling operations within graph neural networks that simultaneously improve effectiveness and reduce computational complexity, i.e., design down-sampling mechanisms that produce compact graph representations while maintaining or enhancing performance and achieving lower computational cost than existing approaches.

Background

Pooling is a key down-sampling operation in graph neural networks that reduces graph size to control model complexity, improve permutation invariance, and enable efficient computation. Common strategies include mean/max/sum pooling, sorting-based pooling, and hierarchical pooling.

Recent methods such as DiffPool learn differentiable cluster assignments but can generate dense graphs after pooling, increasing computational complexity to O(n^2). More efficient and effective pooling mechanisms are needed to balance representational quality with scalability.