Add SpaPooling: hierarchical graph pooling operator

[rodrigueg]

Summary

Adds SpaPooling (Soft Partition Assignment Pooling), a hierarchical graph pooling operator published at the International Conference on Big Data Analytics and Knowledge Discovery (DaWaK 2025).

"SpaPool: Soft Partition Assignment Pooling for Graph Neural Networks"
Govan, R., Scherrer, R., Fournier-Viger, P., Selmaoui-Folcher, N., DaWaK 2025
https://doi.org/10.1007/978-3-032-02215-8_27

Method

SpaPooling selects a subset of representative nodes via a configurable selection module (TopK or SAGPool), then computes a soft assignment matrix S between all nodes and the representatives. The coarsened graph is obtained as:

• O = Sᵀ X  (pooled node features)
• A' = Sᵀ A S  (pooled adjacency)

Three association modes are supported: scalar dot-product, cosine similarity, and scaled dot-product attention. Several regularization losses are available: DiffPool, MinCut, DMoN, or a custom combination.

Checklist

• reset_parameters() implemented
• __repr__() implemented
• Docstring references the paper
• Tests cover shapes, batching, all select/association/loss modes, and gradient flow
• flake8 passes