Global to Local: Topology-Preserving Adaptive Graph Pooling via Granular-Ball
Sen Zhao, Gaojie Xu, Shuyin Xia, Yifan Guan, Yi Liu, Yi Wang, Wei Wang
cs.AI
Quick verdict
TPAGP pools graphs global-to-local via granular balls; it tops 6 benchmarks, but its DD '>3%' claim contradicts Table 2 and statistics/seeds are unreported.
Summary
The paper proposes TPAGP, a graph-pooling method that inverts the usual bottom-up coarsening of hierarchical pooling: rather than starting from individual nodes, the whole graph is treated as the coarsest granular ball and recursively refined. The graph is first split into √N balls centered on the highest-degree nodes via multi-source BFS; each ball is then recursively halved (two highest-degree centers per split) whenever a quality score Q — the summed durations of persistence bars from a Rips-complex filtration over node features — indicates the children beat the parent. The resulting ball tree drives a multi-granularity network alternating graph convolution and pooling, ending in multiset readout and an MLP classifier. On six TUDataset benchmarks (MUTAG, MSRC_9, BZR, DD, PTC_MR, IMDB-MULTI), TPAGP reports the best accuracy on all six (average 82.94% vs 77.80% for the strongest baseline, GINICL), including 97.56% on BZR and 76.47% on PTC_MR; ablations show both granular-ball pooling and the multi-granularity network contribute. Key limitations: the claimed '>3%' margins on DD and PTC_MR conflict with its own Table 2 (DD is +1.95 over SUGAR); p<0.01 significance is asserted without test details, run counts, or standard deviations; no complexity or runtime analysis of the persistent-homology quality gate; and the hyperparameter appendix reference is broken ('??').
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
4 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
- Prior hierarchical pooling methods operate bottom-up from individual nodes, missing global-to-local patterns and per-subdomain adaptive granularity (Sec. 1).
- TPAGP adaptively partitions a graph into granular balls using degree-based centers, multi-source BFS, and a persistent-homology quality gate on recursive splits (Secs. 3.1-3.2).
- TPAGP achieves the highest test accuracy on all six benchmarks (avg 82.94%, Table 2), with improvements claimed significant at p<0.01.
- Both granular-ball pooling and the multi-granularity network are necessary, per the Table 3 ablations (e.g., BZR 97.56 → 92.68 → 85.91).
Strengths
Genuinely inverted pooling paradigm
Sec. 1 and Fig. 1(c) reframe pooling as global-to-local refinement starting from the whole graph as one granular ball, in contrast to the node-dropping (TopKPool, SAGPool) and node-clustering (DiffPool, MinCutPool) baselines that coarsen bottom-up — a fresh angle relative to the compared literature.
Concrete, implementable refinement algorithm
Sec. 3.1 fully specifies the mechanism: degree-ranked center selection (Eq. 1), multi-source BFS assignment by shortest-path competition (Eq. 2), and a quality-gated binary split rule (Eqs. 3-4) with an explicit termination condition (Eq. 4).
Interpretable adaptive-granularity criterion
Broad empirical sweep with component ablation
Weaknesses
Internal inconsistency on a headline claim
Unverifiable statistical claim
No cost or scalability analysis
Undefined edge-case behavior
Presentation quality artifacts undermine rigor
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper (arXiv 2609.04978, cs.AI, marked 'Preprint. Under review.') proposes TPAGP, a graph-pooling method for graph classification that reverses the standard bottom-up coarsening of node-dropping (TopKPool, SAGPool) and node-clustering (DiffPool, MinCutPool) methods. Central claims: (1) pooling should proceed global-to-local, treating the whole graph as the coarsest granular-ball and recursively refining until each subdomain reaches adaptive granularity (Sec. 1, Fig. 1c); (2) a persistent-homology quality score can gate recursive splits (Secs. 3.1-3.2); (3) TPAGP achieves the highest test accuracy on all six benchmarks with p<0.01 significance (Table 2). Sec. 4.3 additionally claims >3% margins over the best alternative on DD and PTC_MR.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 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 TPAGP and what problem does it address?
TPAGP (Topology-Preserving Adaptive Graph Pooling) is a graph neural network pooling method for graph classification. Unlike standard bottom-up pooling — which drops or merges individual nodes step by step — it starts from the entire graph as one coarse 'granular ball' and recursively splits it into smaller topologically coherent regions, letting each region settle at its own granularity. The goal is to preserve both global and local structure during graph compression.
How well does TPAGP perform compared to other methods?
The paper reports the highest test accuracy among 14 compared methods on all six TUDataset benchmarks: MUTAG 89.47%, MSRC_9 97.78%, BZR 97.56%, DD 83.05%, PTC_MR 76.47%, IMDB-MULTI 53.33% — averaging 82.94% versus 77.80% for the best prior baseline (GINICL). Independent verification is advisable: the stated >3% gain on DD does not match its own table, and no variances or test details accompany the p<0.01 significance claim.
Is TPAGP's code and data available?
The paper links an anonymous code repository (anonymous.4open.science/r/TPAGP), and all six datasets come from the public TUDataset collection. However, per-dataset hyperparameters are missing — the text cites a broken 'appendix ??' reference — so reproducing the exact numbers may require re-tuning.
What is a granular ball in this context?
A granular ball is a cluster-like region from granular-ball computing: a group of data points treated as one unit at a chosen level of granularity. In TPAGP, balls are sets of graph nodes that are topologically cohesive — assigned by multi-source BFS from high-degree hub nodes — and the quality of each ball is scored using persistent homology (persistence-bar durations from a Rips complex built on node features), which decides whether a ball should split further.
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 1-4 (including the Fig. 3 hyper-parameter curves and Fig. 4 pooling visualizations) are referenced but their graphical content is not part of the extracted text.
- No hyperparameter table appears anywhere in the provided text; Sec. 4.2 cites 'appendix ??' and the text ends after Appendix A's dataset introductions.
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.