Research map / Graph learning / GNN architectures
SkipPool: Improved Sparse Hierarchical Graph Pooling with Differentiable Exploration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SKIPPOOL, a sparse hierarchical graph pooling operator that uses Gumbel-softmax-based learnable skipping on top of Top-K scoring to avoid over-representing regions, evaluated on graph classification and point-cloud reconstruction.
Problem
Standard sparse top-K graph pooling (SSP) over-represents symmetric regions of the graph signal space and leaves other regions unpooled, losing graph-level information for graph classification.
Approach
Keep SAGPool-style node scoring but couple Top-K ranking with a differentiable skipping mechanism: Gumbel-softmax-sampled distance thresholds group nearby redundant nodes and skip to distant, unrepresented nodes, yielding more balanced pooling while staying sparse.
Key results
- SKIPPOOL reaches 77.18% on D&D and 73.67% on NCI1, outperforming ASAPPOOL, TOPKPOOL and SAGPOOL on most hierarchical-pooling benchmarks (fixed ratio 0.5).
- SKIPPOOL-FULL achieves state-of-the-art on REDDIT-BINARY (87.72% vs 87.57% for KMIS max) in the non-ratio-specified setting.
- In the GAE point-cloud reconstruction test, SKIPPOOL-FULL attains the best average rank (1.67) among parameterized baselines, with lower MSE than SAGPOOL and TOPKPOOL on most graphs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace ASAPPOOL with SKIPPOOL | GCN graph-classification model · D&D | accuracy: 76.87±0.7 -> 77.18±0.48 | Helped |
| Replace ASAPPOOL with SKIPPOOL | GCN graph-classification model · PROTEINS | accuracy: 74.19±0.79 -> 74.76±0.47 | Helped |
| Replace ASAPPOOL with SKIPPOOL | GCN graph-classification model · NCI1 | accuracy: 71.48±0.42 -> 73.67±0.51 | Helped |
| Replace ASAPPOOL with SKIPPOOL | GCN graph-classification model · NCI109 | accuracy: 70.07±0.55 -> 72.23±0.73 | Helped |
| Replace ASAPPOOL with SKIPPOOL | GCN graph-classification model · FRANKENSTEIN | accuracy: 66.26±0.47 -> 62.33±0.61 | Hurt |
| Use SKIPPOOL-FULL instead of KMISmax | GCN graph-classification model · D&D | accuracy: 76.91±1.06 -> 76.66±0.94 | Hurt |
| Use SKIPPOOL-FULL instead of KMISmax | GCN graph-classification model · REDDIT-BINARY | accuracy: 87.57±1.96 -> 87.72±1.66 | Helped |
Surprising findings
- Score values that deviate from the ideal gating values 0 and 1 can still perform well or outperform, challenging the gating interpretation of SSP scores.
- ASAPPOOL's FRANKENSTEIN advantage comes from becoming dense, producing near-complete graphs at the last layer for about 88% of the dataset.
Limitations
- Skipping is CPU-bound sequential index operations, so runtime increases relative to pure Top-K methods (1.4x-1.7x on D&D vs ASAPPool).
- On FRANKENSTEIN, SKIPPOOL underperforms ASAPPOOL substantially (62.33 vs 66.26), which the authors attribute to ASAPPool's denser outputs.
- SKIPPOOL-FULL does not allow direct pooling ratio specification; ratio is controlled indirectly via stride length F.
Experimental setup
Datasets: PROTEINS D&D NCI1 NCI109 FRANKENSTEIN REDDIT-BINARY Grid2D Ring Airplane Car Guitar Person
Benchmarks: PROTEINS D&D NCI1 NCI109 FRANKENSTEIN REDDIT-BINARY Grid2D Ring Airplane Car Guitar Person
Metrics: accuracy reconstruction MSE runtime per epoch
Baselines: SET2SET GLOBAL-ATTENTION SORTPOOL DIFFPOOL TOPKPOOL SAGPOOL ASAPPOOL BDO GRACLUS EDGEPOOL KMIS
Builds on: SAGPool TopKPool KMIS GIN Gumbel-softmax
Closest papers by meaning
- Geometry-Aware Edge Pooling for Graph Neural Networks (NeurIPS 2025 · similarity 0.90)
- Enhancing Graph Classification Robustness with Singular Pooling (NeurIPS 2025 · similarity 0.89)
- MaxCutPool: differentiable feature-aware Maxcut for pooling in graph neural networks (ICLR 2025 · similarity 0.89)
- Graph Parsing Networks (ICLR 2024 · similarity 0.88)
- ENAHPool: The Edge-Node Attention-based Hierarchical Pooling for Graph Neural Networks (ICML 2025 · similarity 0.88)
- Boosting Graph Pooling with Persistent Homology (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.