Hierarchical Possession-Aware Graph Pointer Network for Pass Receiver Selection
Jingyi Wang, Da Li, Kaixin Wang, Zhangqin Huang
cs.AI · cs.SI
Quick verdict
HPGPN beats XGBoost by 3.31pp on six StatsBomb datasets via graph+history modeling; consistent gains, but no significance tests or code release.
Summary
This paper proposes HPGPN, a Hierarchical Possession-Aware Graph Pointer Network for pass receiver selection in football under a broadcast-like partial-observation setting: event-centered StatsBomb 360 freeze frames expose only visible players without stable identities, so the task becomes selecting from a variable-size set of anonymous visible teammates. The model combines a GraphSAGE encoder over a passer-candidate-defender interaction graph, a fixed event-context encoder covering current attributes plus the two preceding actions, a dual-branch Transformer-based possession-history encoder (all-event and pass-only branches with attention pooling), hierarchical candidate refinement via gated residual fusion, and a two-stage glimpse pointer head. On six public competitions (Bundesliga 23/24, FIFA World Cups 2022 and 2023, Euros 2020 and 2024, Women's Euro 2025), HPGPN is best on all six, raising average accuracy from 0.5387 (XGBoost) to 0.5718 (+3.31 percentage points); ablations on Euro 2024 and Women's Euro 2025 confirm each component, and K=40 maximizes the history-window benefit. Cross-dataset transfer without fine-tuning degrades asymmetrically, with men-to-women transfer holding up better than women-to-men. Main limitations: no statistical significance testing, confidence intervals, or multi-seed variance despite small gaps (e.g., 0.6001 vs 0.5986 in the head comparison); unclear whether baselines received comparable history features; potential selection bias from filtering to passes whose receiver is visible; no code release; ablations restricted to two of six datasets.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 3
● low 1
4 strengths · 4 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
- HPGPN achieves the best pass receiver selection accuracy on all six evaluated StatsBomb competitions, improving average accuracy from 0.5387 (XGBoost) to 0.5718 (+3.31pp).
- Graph-based interaction modeling, fixed event context, and the dual-branch possession history each measurably contribute, per ablations on Euro 2024 and WEuro 2025 (Tables 3-5).
- HPGPN transfers across competitions without fine-tuning without collapsing, with an asymmetric pattern: men-to-women transfer holds while women-to-men drops up to 0.0483 accuracy (Fig. 2).
- A history window of K=40 preceding possession events is optimal, balancing contextual benefit against historical noise (Fig. 3).
Strengths
Practical partial-observation formulation
Task cast as variable-size selection over anonymous visible teammates (Sec. 3) matches real broadcast-style data constraints and scales beyond fully-tracked or identity-annotated settings.
Consistent, quantified wins across six competitions
Table 2: HPGPN best on all six datasets, average 0.5718 vs XGBoost 0.5387 (+3.31pp), with per-dataset gains from +1.83pp (BL 23/24) to +5.55pp (Euro 2020); SoccerMap trails at 0.4077.
Unusually thorough ablation grid
Analyses beyond headline accuracy
Weaknesses
No uncertainty quantification
Baseline input parity unclear
Undocumented visibility filtering
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This paper studies pass receiver selection in the "un-xPass" setting: given a pass event with a StatsBomb 360 event-centered freeze frame, predict the intended receiver among visible teammates only — no full trajectories, no stable player identities, and a candidate set whose size varies per pass (Secs. 1 and 3). Roughly 182,000 visible-pass samples across six competitions (Table 1) support the study. The authors claim three things: (1) HPGPN achieves the best accuracy on all six datasets; (2) every architectural ingredient — the interaction graph, fixed event context, dual-branch possession history, and glimpse pointer head — contributes measurably; (3) the model transfers across competitions without fine-tuning. The scope is honestly bounded: inputs are restricted to pre-pass information, and the task is a standalone prediction problem rather than a valuation metric.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 4 weaknesses, analyzed in detail
- 4 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 HPGPN and what problem does it solve?
HPGPN (Hierarchical Possession-Aware Graph Pointer Network) is a deep learning model that predicts which teammate a football player will pass to, using only event-centered freeze-frame snapshots (StatsBomb 360) where some players are invisible and unnamed — mimicking what broadcast footage reveals, rather than requiring full player-tracking data or stable player identities.
What data was used and is it public?
Public StatsBomb Open Data with 360 freeze-frame annotations, covering six competitions: Bundesliga 2023/24, FIFA World Cup 2022, Euro 2020, Euro 2024, Women's Euro 2025, and Women's World Cup 2023 — roughly 182,000 visible-pass samples in total. The data can be downloaded free from StatsBomb's open-data GitHub repository.
How much better is HPGPN than previous methods?
On average across the six datasets, accuracy rises from 0.5387 for the strongest prior baseline (an XGBoost ranker on handcrafted spatial features) to 0.5718 — a 3.31 percentage-point absolute gain — with HPGPN best on every dataset; SoccerMap reached 0.4077 and a closest-teammate heuristic 0.3783.
Is the code available?
The paper does not mention a public code release. The underlying data is freely available, but reimplementation would require reconstructing the feature engineering from the paper's equations, and some details (exact feature lists, layer counts) are not fully specified.
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.
- Values inside Figs. 2-4 are extracted as jumbled inline text (axis labels, legend entries, delta lists); individual cross-dataset delta cells in Fig. 2 and exact curve values in Figs. 3-4 cannot be reliably mapped to specific points, so those analyses were judged qualitatively from the surrounding prose plus the clearly attributable numbers (e.g., K=40 test accuracies 0.6001/0.5675).
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.