Geometric Flow enhanced Graph Coarsening
Chaoqun Fei, Guoxuan Li, Tinglve Zhou, Chuanqing Wang, Yangyang Li
cs.AI
Quick verdict
RicciPool adds Ollivier-Ricci flow edge-reweighting to graph pooling; thorough ablations, but wins lack significance tests and reported numbers conflict.
Summary
The paper proposes RicciPool, a clustering-based graph pooling method for graph classification that injects discrete geometry into coarsening. An Ollivier-Ricci flow (Algorithm 1) iteratively reweights edge weights — contracting positive-curvature, clique-like edges and stretching negative-curvature, tree-like edges — treating weights as distances; the reweighted matrix then feeds a MinCutPool-style spectral objective (Eq. 13) learned by a GCNN to produce a soft cluster assignment, followed by standard coarsening (Eq. 14). Curvature computation is offline preprocessing. On seven benchmarks (ENZYMES, PROTEINS, D&D, IMDB-B, IMDB-M, COLLAB, REDDIT-M-5K) against 13 baselines (K=32, l=5, α=0.5, η=1.0, 10-fold CV), RicciPool reports the best accuracy on D&D (82.99), PROTEINS (79.46), IMDB-B (76.00) and IMDB-M (51.62), while losing on ENZYMES to SpreadPool/MagPool (62.42 vs 66.30/67.50) and on COLLAB to SEP (75.30 vs 81.40). Ablations cover flow iterations, pooling ratio, η/α, stopping threshold, SBM-based interpretability, and scalability, and the paper honestly flags dense (COLLAB) and very sparse (REDDIT-M-5K) graphs as weak spots. Main limitations: no statistical significance tests despite narrow margins (IMDB-M only +0.09 over SEP), O(N⁴log²N) per-iteration curvature cost (REDDIT-M-5K training ~140 hours), and internal inconsistencies between prose and Table VI.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Ollivier-Ricci curvature flow reweighting captures higher-order neighborhood connectivity that existing clustering-based pooling methods ignore, and feeds better cluster assignments.
- With fixed settings (K=32, l=5, α=0.5, η=1.0), RicciPool achieves the best average accuracy among 13 compared methods on D&D (82.99), PROTEINS (79.46), IMDB-B (76.00) and IMDB-M (51.62).
- Curvature sign is a usable interpretability signal: as clusters separate, intra-cluster edges become positive-curvature dominant and inter-cluster edges negative-curvature dominant (SBM simulation, Fig. 10).
- Decoupling curvature computation as offline preprocessing leaves GCNN training and inference unchanged, with the extra cost manageable except on dense graphs.
Strengths
Comprehensive, current baseline coverage
Table II compares 13 methods across three categories on seven benchmarks under a shared GCNN framework, including recent geometry-aware baselines: ORC (TMLR 2024), SpreadPool and MagPool (NeurIPS 2025), plus CCP-GNN (TNNLS 2025) and MID.
Unusually deep ablation and convergence analysis
Beyond iteration count (Table IV), the paper links curvature-flow convergence speed (Table V: ~10 iterations for PROTEINS, ~6 for IMDB-B) to the optimal l (l=3 vs l=2), and varies pooling ratio (Fig. 4), η and α (Fig. 5), stopping threshold δ (Fig. 12), with stability curves in Appendix A.
Honest failure-mode and suitability analysis
Mechanistic interpretability evidence
Reproducibility aids
Weaknesses
No statistical significance testing
Internal numeric inconsistencies
Prohibitive asymptotic and observed cost
Loose theory presented as equivalence
Baseline fairness and protocol ambiguity
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Review: Geometric Flow enhanced Graph Coarsening (RicciPool)
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is RicciPool in graph neural networks?
RicciPool is a graph pooling (coarsening) method for graph classification. It first runs an Ollivier-Ricci curvature flow that reweights edge weights — shrinking clique-like, positive-curvature edges and stretching tree-like, negative-curvature ones — then learns a spectral-clustering-style soft assignment matrix (in the spirit of MinCutPool) with a GCNN, and merges nodes within each cluster into supernodes. Curvature is computed offline before training.
How is RicciPool different from MinCutPool?
RicciPool keeps MinCutPool's relaxed minCUT spectral loss and coarsening step, but applies them to a graph whose edge weights were pre-evolved by an Ollivier-Ricci flow, so densely connected neighborhoods get short distances and sparse or bridge structures long ones. The paper reports consistent gains over MinCutPool on all seven datasets, e.g., 82.99 vs 78.89 on D&D and 79.46 vs 76.85 on PROTEINS.
What are RicciPool's main limitations?
Three stand out. Cost: Ollivier-Ricci curvature costs O(N⁴log²N) per flow iteration, so dense graphs are expensive — on COLLAB, precomputation (24,803-37,287 s) exceeds training time, and REDDIT-M-5K training takes about 140 hours. Applicability: the curvature contrast the method exploits vanishes on very dense (COLLAB) or near-tree (REDDIT-M-5K) graphs. Evidence quality: no statistical significance tests are reported, and the best hyperparameters (η, iteration count l) vary by dataset.
Where does RicciPool beat other graph pooling methods, and where does it lose?
Per Table II, it reports the best accuracy among 13 methods on D&D (82.99), PROTEINS (79.46), IMDB-B (76.00) and IMDB-M (51.62). It loses to SpreadPool/MagPool on ENZYMES (62.42 vs 66.30/67.50), to SEP on COLLAB (75.30 vs 81.40), and to MID on REDDIT-M-5K (55.20 vs 56.04).
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 3-14 were extracted with garbled axes and legends; figure-based quantitative claims rely on captions and prose only.
- Some equations (e.g., Eq. 6) render garbled in the extraction, so judgments about signs/typos in formulas carry some uncertainty.
- The extraction appears near-complete (all sections and Tables I-VI legible), but subscripts in equations are occasionally unreliable.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.