Differentiable Pooling (DiffPool)

DiffPool learns a soft assignment matrix S ∈ R^(n_l × n_(l+1)) at each layer, mapping nodes from the current graph to clusters in the next coarser graph. Each entry S_ij represents the probability of node i being assigned to cluster j. This differentiable clustering allows gradients to flow through the pooling operation, enabling end-to-end training of the entire architecture for tasks like graph classification.

By stacking multiple DiffPool layers, the model builds a hierarchical representation of the input graph. Each layer reduces the number of nodes while learning increasingly abstract features:

• Layer 1: Captures local substructures and functional groups
• Layer 2: Identifies motifs and ring systems
• Layer 3: Encodes global molecular topology This mirrors the natural hierarchy in molecular graphs, from atoms to functional groups to the full molecule.

Beyond node features, DiffPool simultaneously learns a coarsened adjacency matrix A_(l+1) = S^T A_l S, which defines the connectivity between clusters in the next layer. This preserves topological information through the pooling hierarchy, allowing the model to learn which substructures are connected. The pooled adjacency acts as a learned link prediction between abstract graph regions, crucial for understanding molecular connectivity patterns.

Unlike earlier pooling methods (sort pooling, top-k pooling) that used non-differentiable operations, DiffPool's soft assignment mechanism is fully differentiable. This means:

• The clustering is learned jointly with the GNN embeddings
• No separate pre-training or hand-crafted pooling heuristics
• The model discovers optimal graph coarsening strategies from data This property is essential for molecular property prediction where the optimal hierarchical decomposition is not known a priori.

DiffPool maintains permutation invariance—the output is independent of node ordering. The assignment matrix S is computed using a GNN that operates on the graph structure, and the pooled features and adjacency are computed via permutation-equivariant operations. This ensures that isomorphic molecular graphs (same atoms, different atom ordering in the input) produce identical pooled representations, a fundamental requirement for molecular machine learning.

To prevent degenerate solutions where all nodes collapse into a single cluster, DiffPool employs two key regularizers:

• Link prediction loss: Encourages nearby nodes to be assigned to the same cluster by minimizing the difference between the original adjacency and the adjacency reconstructed from assignments
• Entropy regularizer: Penalizes assignment matrices with low entropy, encouraging clusters to be used uniformly rather than collapsing These ensure meaningful, interpretable cluster hierarchies emerge during training.

A classical graph similarity metric quantifying the minimum cost of edit operations (node/edge insertion, deletion, substitution) to transform one graph into another. DiffPool embeddings are often trained to approximate GED for similarity search in molecular databases. Computing exact GED is NP-hard, making learned neural proxies like those from DiffPool essential for scalable virtual screening.