Source-linked AI summary
Hierarchical Possession-Aware Graph Pointer Network for Pass Receiver Selection
Jingyi Wang, Da Li, Kaixin Wang, Zhangqin Huang
TL;DR
Pass receiver selection under partial, variable visibility requires modeling candidate interactions and recent possession context rather than relying only on a freeze frame. HPGPN combines graph, event-context, and dual-branch history encoders with hierarchical refinement and pointer-based candidate scoring. It consistently improves performance across public datasets, while ablations support the value of its principal components.
Problem
Pass receiver selection from event-centered freeze frames must identify the intended teammate among partial, variable visible candidates while accounting for spatial interactions and possession history.
Method
HPGPN jointly models the current pass graph, fixed event context, and dual-branch dynamic possession history, then hierarchically refines candidates and scores them with a glimpse pointer head.
Results
HPGPN consistently outperforms heuristic, spatial deep learning, and feature-based baselines across multiple public datasets.
Takeaways & Limitations
Receiver selection benefits from structured modeling of player interactions, event-level context, and possession-level temporal information.
Takeaways & Limitations
The formulation uses only information available before or at the pass moment and excludes the true receiver, pass outcome, post-pass movement, and subsequent events.
Abstract
from arXiv · showhide
Pass receiver selection is a fundamental task in football analytics, aiming to predict the intended receiver under a given game state. This task is challenging with event-centered freeze-frame observations, a broadcast-like setting that provides only partial and variable player visibility without complete trajectories or stable player identities. The model must therefore reason over anonymous visible candidates, opponent pressure, and recent context under partial observation. To address this setting, we propose a Hierarchical Possession-aware Graph Pointer Network (HPGPN), which formulates pass receiver selection as variable-size candidate prediction over visible teammates. HPGPN jointly models current player interactions, local event context, and possession-level temporal dynamics. It represents the current pass situation with a graph, incorporates fixed event context, and uses dynamic possession history to capture how the attacking sequence evolves. Candidate representations are refined hierarchically by integrating spatial, contextual, and historical evidence, and a glimpse pointer head scores the receiver candidates. Experiments on public football event and freeze-frame data show that HPGPN improves pass receiver selection performance. Ablation studies demonstrate the effectiveness of graph-based interaction modeling, fixed event context, and dual-branch dynamic possession-history modeling.
1 Introduction
Pass receiver selection requires reasoning over multiple visible teammates, spatial and tactical context, and preceding possession development under partial observation. HPGPN addresses this with hierarchical graph, context, and history modeling for variable-size receiver prediction.
- Pass receiver selection asks which teammate the ball carrier intends to receive the pass given the current game state.
- Event-centered freeze frames provide partial player snapshots without dense trajectories, creating a broadcast-like prediction setting.
- Current spatial layout alone is insufficient because receiver choice reflects preceding actions, passing patterns, attacking direction, and possession evolution.
- HPGPN combines a current player-interaction graph, fixed event context, and dual-branch possession history in a hierarchical candidate-refinement architecture.
- A glimpse pointer head scores the available receiver candidates directly, supporting variable-size visible candidate sets without a fixed global player vocabulary.
2 Related Work
Prior pass-receiver research progressed from hand-crafted spatial features toward tracking, video, multimodal, spatial-map, and graph-based approaches. Pointer and graph networks provide the architectural basis for selecting candidates from variable relational inputs.
- Early pass prediction methods used hand-crafted distances and angles among the passer, candidate receivers, and nearby opponents.
- Later approaches incorporated tracking data, video frames, and multimodal inputs to model movement, defensive organization, and visual context.
- SoccerMap used pitch-level spatial maps, while un-xPass estimated pass receiver selection as part of a pass creativity metric.
- Pointer Networks select elements from input sequences when output size depends on the input, and graph neural networks encode relational structure.
- Graph Pointer Networks combine graph embedding with pointer decoding to capture input-node relationships before selecting an element.
3 Problem Formulation
The task predicts the intended receiver from visible teammates in each pass-specific freeze frame, whose candidate set varies across events. Predictions use only information available by the pass moment and select the highest-probability current candidate.
- Pass receiver selection is formulated as candidate selection over a variable-size set of visible teammates rather than classification over a fixed player vocabulary.
- Each event-centered freeze frame contains the passer, visible teammates, visible opponents, and their spatial information.
- The ground-truth receiver is represented as candidate cn among Np visible teammate candidates, excluding the passer.
- The model estimates P(yp = n | p, Cp) over n ∈{1, . . . , Np}, using the current pass graph, fixed event context, and dynamic possession history.
- Inputs are restricted to information available before or at the pass, excluding the true receiver, pass outcome, post-pass movement, and subsequent events.
- The predicted receiver is the candidate with maximum estimated probability, while possession history is bounded by a maximum window length.
4 Methodology
HPGPN hierarchically refines candidate and decision-query representations by combining current player relations, fixed event-level cues, and recent possession context. Its graph encoder explicitly represents passing options and defensive interactions before later context and refinement modules.
- HPGPN uses a candidate-centric architecture that progressively refines receiver representations and the pass-level decision query.
- Current pass graph: The current pass is modeled as a pass-centric graph of the passer, candidate receivers, and defenders to represent passing options and defensive interactions.
- Current pass graph: Directed passer-to-candidate edges and bidirectional candidate-to-defender edges encode the graph’s interaction structure.
- Current pass graph: Node features combine spatial and node-type information, while candidate nodes additionally include pass distance, pass angle, and defensive proximity.
- Current pass graph: GraphSAGE encodes the nodes, and pooled graph representations summarize the current spatial decision context for subsequent context encoding and candidate refinement.
- Fixed event context: The fixed event context encoder captures current pass attributes, freeze-frame summaries, and gamestate information from the two immediately preceding actions.
4.3 Dual-Branch Dynamic Possession History Encoder
The encoder models preceding possession events with time-aware tokens and a temporal Transformer, then derives complementary all-event and pass-only history contexts.
- Dual-branch representation: The dual-branch encoder captures complementary evidence from overall possession development and historical passing patterns.
- History construction: The encoder retains up to K preceding events from the same possession, ordered from recent to early events and excluding the current pass.Each event includes action, spatial, pressure, outcome, and relative-time information.
- History construction: Each historical event becomes a time-aware token by combining event features with temporal features.
- Temporal contextualization: A temporal Transformer contextualizes the historical token sequence before downstream history aggregation.
- Dual-branch representation: The shared contextualized sequence supports both all-event and pass-only history contexts.
All-Event History Context
The all-event branch summarizes overall possession development by attending over all contextualized history tokens with a pass-conditioned query.
- Query construction: The branch constructs a pass-conditioned query from the passer, graph, fixed context, and history-availability features.
- History attention: Attention scores are computed over the contextualized history tokens and normalized across the retained history.
- History aggregation: The resulting weighted token aggregation forms the all-event history context.
- History aggregation: When no history is available, the all-event context is set to zero and a history-absence indicator is retained.
Pass-Only History Context
The pass-only branch isolates historical pass actions while preserving surrounding possession context, then uses them to refine each candidate representation.
- Pass-history encoding: The pass-only branch summarizes historical passes that may indicate recent passing patterns and attacking direction.
- Pass-history attention: A pass-history query incorporates the passer, graph, fixed context, and availability and length features for both history branches.
- Pass-history attention: If no historical passes exist, the pass-history context is set to zero and its absence is indicated.
- Candidate refinement: Hierarchical refinement integrates graph-based spatial relations, fixed event context, and candidate-specific pass-history evidence into each final candidate representation.
- Candidate refinement: The pass-history fusion uses a learnable scalar initialized to zero.
4.5 Glimpse Pointer Receiver Selection Head
The glimpse pointer head converts refined, variable-size candidate representations into a receiver distribution by updating a pass-level query through candidate attention.
- Pointer scoring: After refinement, each candidate is represented by an embedding suitable for scoring a variable-sized candidate set.
- Pointer scoring: An initial pointer query is constructed from the pass-level context.
- Pointer scoring: A glimpse attention step summarizes the refined candidate embeddings and updates the pointer query.
- Pointer scoring: The updated query produces a pointer logit for every candidate.
- Pointer scoring: Candidate logits are normalized over the current candidate set to obtain the receiver distribution.
4.6 Training Objective
The model is trained to maximize the likelihood of the ground-truth receiver among each pass's candidate set, with accuracy used for evaluation. Per-pass inference combines graph message passing, temporal self-attention, candidate–history attention, and feature projections within a stated complexity bound.
- The training objective uses candidate-level negative log-likelihood for the ground-truth receiver.The objective is defined over each pass and its candidate set.
- Accuracy is the primary evaluation metric.
- Per-pass inference complexity accounts for graph message passing, temporal self-attention, candidate–history attention, feature projections, fusion layers, and pointer-based scoring.The terms are respectively associated with graph interactions, historical-event attention, candidate–history interactions, and projection or scoring operations.
- The complexity depends on receiver candidates, visible opponents, historical events, and hidden dimension, denoted by N, M, L, and d.The graph is described as small, and the history length satisfies L ≤ K, keeping inference manageable.
5 Experiments
Experiments evaluate HPGPN on public StatsBomb 360 data against several baselines and through transfer, history-window, candidate-set, and component analyses. HPGPN achieves the strongest reported receiver-selection accuracy while benefiting from possession history, event context, GraphSAGE encoding, and glimpse-based candidate comparison.
- Experimental setup: The experiments use public StatsBomb 360 freeze frames, where only camera-visible players are available, and evaluate receiver-selection accuracy among visible candidates.Matches are split into held-out test matches and training/validation samples; all methods use the same input constraints as un-xPass.
- Comparison with baselines: HPGPN achieves the best accuracy on all six datasets, improving average accuracy over XGBoost from 0.5387 to 0.5718.The absolute gain is 3.31 percentage points; the largest gains are 5.55 points on Euro 2020 and 4.42 points on WEuro 2025.
- Cross-dataset transferability: Cross-dataset transfer does not collapse, with effective transfer within men’s competitions and from men’s to women’s competitions, but larger drops in the reverse direction.The transfer analysis compares source-to-target accuracy with the target in-domain baseline.
- Sensitivity to maximum history window length: K = 40 gives the best test accuracy on both Euro 2024 and WEuro 2025, reaching 0.6001 and 0.5675, respectively.Smaller windows provide insufficient context, whereas larger windows may add less relevant earlier events.
- Performance under different candidate-set sizes: Performance decreases as candidate sets grow, while Top-2/Top-3 accuracy and MRR remain strong across common sizes dominated by six-to-nine-candidate samples.The analysis indicates that HPGPN can provide useful rankings even when Top-1 selection is incorrect.
- Ablation study: Ablations show that fixed event context, fixed short-term history, dynamic possession history, and its contextualized metadata improve performance; GraphSAGE and the glimpse pointer head are the strongest tested choices.The glimpse pointer head outperforms independent MLP scoring, while GraphSAGE outperforms GCN and GAT.
6 Conclusion
HPGPN addresses pass receiver selection under partial observation by jointly modeling current player interactions, event context, and possession history. Across public datasets, it outperforms several baseline types, while ablations support its key components.
- HPGPN jointly models the current pass graph, fixed event context, and dynamic possession history.
- The model selects an intended receiver from a variable set of visible candidate teammates using a glimpse pointer head.
- Experiments on multiple public datasets show consistent improvements over heuristic, spatial deep learning, and feature-based baselines.
- Ablation studies validate the contribution of HPGPN’s key components.
- The results support structured modeling of player interactions, event-level context, and possession-level temporal information for receiver selection.