DiffPool

Early GNNs were flat: they learned node embeddings and then collapsed them with a single global readout, throwing away the hierarchical structure that makes graphs interesting. DiffPool learns the hierarchy instead. At each layer a GNN predicts a soft assignment matrix mapping nodes to clusters; nodes are pooled into cluster embeddings and the coarsened graph becomes the input to the next layer. Two auxiliary losses keep the assignments sensible: a link-prediction loss encouraging nearby nodes to cluster together, and an entropy penalty pushing assignments toward crisp decisions. The whole thing is differentiable, so pooling is learned jointly with the classification objective rather than fixed upfront.\n\nThe NeurIPS 2018 paper reports that adding DiffPool to a Structure2Vec base lifts ENZYMES accuracy from 61.10 to 63.33 and D&D from 78.92 to 82.07, with average gains of 5-10% across graph-classification benchmarks and a new state of the art on four of five datasets. A practical bonus: GraphSAGE with DiffPool trained 12x faster than the same model with Set2Set pooling, because coarsening shrinks the graphs the convolutions operate on. DiffPool established hierarchical pooling as a first-class GNN component and inspired a long line of successors.

Paper-reported: S2V + 2 DiffPool layers reaches 63.33 on ENZYMES and 82.07 on D&D (vs 61.10/78.92 base); average 5-10% gains across benchmarks, new SOTA on 4 of 5 datasets (Ying et al., NeurIPS 2018).