PaperlayerView exact revision on arXiv

Source-linked AI summary

Automated Detection of Match Phases in Football from Spatio-Temporal Tracking Data Using Graph Neural Networks

Vincent Renner, Nils Koster, Pascal Bauer, Melanie Schienle

arXiv:2610.11571v1stat.APstat.ML

TL;DR

The paper addresses open questions about graph modeling of football’s spatial and temporal tracking data and phase classification. It combines player-level graphs with a sequence model; the GNN-LSTM outperforms the tabular baselines, with the strongest baseline scoring 4.6% lower in macro F1.

  • Problem

    How GNNs compare with tabular baselines and how graph topology and message-passing choices affect football phase classification remain open questions.

    Source: §1 · p. 1

  • Method

    The paper uses player-level graph representations and a sequential GNN approach to identify both teams’ match phases at one-second resolution.

    Source: §1 · p. 2

  • Results

    4.6% lower macro F1 for the strongest baseline, XGBoost on standard features, than the GNN-LSTM, which outperformed all tabular baselines.

    Source: §6 · p. 11

  • Takeaways & Limitations

    The framework automates phase labeling and can be retrained on phase definitions used by professional teams.

    Source: §6 · p. 12

    There is no separate test set, so absolute performance figures may be slightly optimistic estimates for unseen matches.

    Source: §4 · p. 7

Abstract

from arXiv · show

Spatio-temporal tracking data has opened new possibilities for detecting complex tactical patterns in football, yet modeling the interactive movements of multiple players remains challenging. This paper proposes a framework combining graph neural networks (GNNs) with a sequential model to classify match phases on a second-by-second basis across a seven-class taxonomy. Match phase classification is tactically meaningful, and the availability of rule-based labels across 203 matches makes it a suitable testbed for a systematic comparison of adjacency constructions and message-passing layers, a question that has received limited attention in existing research. Our selected GNN-LSTM model outperforms all aggregated-feature baselines, including XGBoost and a Long Short-Term Memory (LSTM) network, as the strongest baseline scores 4.6% lower in macro F1. Graph representations using a domain-informed Delaunay triangulation that approximates passing lanes, paired with a custom Spatial Edge-Augmented Convolution (SEAConv) layer that injects edge attributes directly into messages, achieve the best performance by capturing spatial dependencies while limiting uninformative messages from redundant edges. Integrated Gradients attributions indicate the model uses spatial player configurations, particularly horizontal positioning, in combination with possession and ball-status indicators. This work offers an automated solution for fine-grained tactical analysis, reducing the need for manual tagging and providing deeper insight into dynamic team behavior.

1 Introduction

The paper addresses open questions about how graph neural networks compare with tabular baselines and how graph topology and message-passing choices affect football phase classification. It proposes player-level graphs and a sequential model for one-second phase identification, with systematic design comparisons and feature attribution.

1 Introduction

  • The paper asks whether GNNs outperform tabular baselines and how topology and message-passing choices affect football phase classification.The authors identify these comparisons as open questions in a tactically meaningful task with tracking-synced labels available at scale.Source: §1 · pp. 1–2
  • The proposed GNN captures spatial and temporal dependencies to identify both teams’ match phases at one-second resolution.It uses player-level features and preserves spatial structure, with a seven-class taxonomy whose three defensive sub-phases are distinguished by player arrangements.Source: §1 · p. 2
  • The study compares domain-driven graph and message-passing choices and uses Integrated Gradients to analyze feature importance.Its contributions also include rule-based labels across 203 matches and an analysis of graph features versus graph-wide possession information.Source: §1 · p. 2
  • Unlike zone-count approaches, the work retains player-level positions and relations, and evaluates graph design choices across 80 runs.The authors contrast their 203-match rule-labeled dataset with related work, noting that this scale comes at the cost of agreement with human annotation.Source: §1 · p. 2

2 Data and Graph Representations

The study uses rule-based phase labels and player-and-ball tracking data from 203 men’s matches. Each timestamp is represented as a graph, with Delaunay triangulation compared against general-purpose adjacency strategies.

2.1 Data Set

  • The dataset comprises 203 men’s matches, with 25 Hz positional data for 22 players and the ball and rule-based phase labels.The matches span the 2019/20–2024/25 seasons and include six named active phases; ‘No Phase’ brings the taxonomy to seven classes.Source: §2.1 · pp. 2–3
  • ‘No Phase’ accounts for 57.2% of match time, while transition phases occupy less time and High Block, Low Block, and Counter Attack total 6.0%.Both teams are labeled each second, and a full match is treated as one sequence.Source: §2.1 · p. 3

2.2 Graph Representations

  • Each timestamp is a directed graph of players and the ball, with node, edge, and global features averaged across 25 frames per second.The 23 nodes carry spatial, movement, and type features; edges encode distance, angle, and relationship type, while global features indicate possession and ball-in-play status.Source: §2.2 · p. 3
  • Delaunay triangulation encodes triangular player arrangements that maintain passing options, addressing a largely unexplored adjacency choice in football.The paper compares it with k-nearest-neighbor, radius, and complete graphs; Figure 1 depicts these strategies on the same observation, with node colors indicating teams and black marking the ball.Source: §2.2 · pp. 3–4, Figure 1 · p. 4

3 Model Design

The model combines message-passing layers and graph pooling with a recurrent sequence model for per-timestep classification. Its layer design incorporates edge information into messages and attention, while the evaluation includes aggregated-feature baselines.

3.1 GNN Layers

  • SEAConv incorporates edge embeddings into both attention coefficients and messages, addressing the limitation that GATConv uses edge attributes only in attention.The layer also treats a node’s own state as a self-loop in the attention softmax and operates on Fourier-encoded coordinates.Source: §3.1 · p. 5
  • Random Fourier features encode spatial coordinates to help resolve fine-grained positional differences.The paper motivates this encoding by the spectral bias of neural networks toward low-frequency functions when using raw low-dimensional coordinates.Source: §3.1 · p. 5

3.2 Model Architecture

  • A GNN block pools each graph into an embedding, concatenates it with embedded global features, and passes the sequence through a unidirectional LSTM.A final classification layer produces predicted probabilities for each class at each timestep; Figure 2 depicts the graph-sequence input and per-timestep outputs.Source: §3.2 · pp. 5–6, Figure 2 · p. 6

3.3 Baseline Models

  • The baselines include logistic regression and XGBoost on aggregated positional features, global-only XGBoost, topological-feature XGBoost, and an LSTM on team aggregates.Tabular models receive ten seconds of lag features, and all models use the same match split.Source: §3.3 · p. 6

4 Empirical Results

The evaluation compares graph constructions and message-passing layers under shared training conditions, then benchmarks the selected GNN against aggregated-feature models and assesses its calibration. The GNN outperforms the baselines, while the evaluation uses validation matches rather than a separate test set.

4 Empirical Results

  • Configurations shared fixed hyperparameters without systematic search; separate tuning might improve absolute scores but confound comparisons.Training used four-part truncated backpropagation through time and 10 epochs of graph-encoder pretraining.Source: §4 · pp. 6–7
  • The 70/30 full-match split has no separate test set, so absolute validation performance may be slightly optimistic; the shared protocol keeps relative comparisons in focus.All GNN results in Tables 2 and 3 are means over four seeds.Source: §4 · p. 7

4.1 Graph Representations and Message-Passing Layers

  • Delaunay significantly exceeds each adjacency alternative on Macro F1; sparse edges prune distant, low-information links while preserving multi-hop relationships.Delaunay was higher in at least 15 of 16 matched pairs in each comparison; the complete graph also had roughly 4.5× more edges and 67% higher average training time than Delaunay.Source: §4.1 · pp. 7–8, Table 2 · p. 7
  • SEAConv significantly exceeds every alternative on Macro F1, including GATConv, with a mean paired difference of +0.0050.The 95% CI is [+0.0042, +0.0060] and Holm-adjusted p < 10^-4; SEAConv leads on every metric except CE loss, where its difference from GATConv is marginal.Source: Table 2 · p. 7, §4.1 · p. 8

4.2 Comparison with Baselines

  • The GNN outperforms every baseline on every metric; the strongest baseline, XGBoost on lagged standard features, trails by 4.6% in macro F1.The table reports results on the validation set and defines percentages relative to the GNN.Source: §4.2 · p. 8, Table 3 · p. 8

4.3 Calibration

  • Most classes are well calibrated, but Low Block and Counter Attack are overconfident in the 0.3–0.8 range; ECE is low across all classes.CE is highest for transition phases and High Block, whose F1 scores also drop; overall CE can hide this because frequent classes dominate it.Source: §4.3 · p. 8, Figure 3 · p. 9

5 Model Interpretability

The interpretability analyses use class-specific Integrated Gradients to compare graph embeddings with global inputs and to rank node features. They find graph embeddings contribute across phases, while horizontal position is especially influential and learned embeddings appear phase-separated in a t-SNE projection.

5 Model Interpretability

  • Integrated Gradients are applied per class to compare graph embeddings with global features, then to assess individual node features while other inputs are held fixed.The analyses use the validation set; node-feature attributions use unconditional feature means as the baseline.Source: §5 · p. 9, §5.1 · p. 9, §5.2 · p. 11

5.1 Importance of Graph Embeddings

  • Graph embeddings receive at least 36.1% of total attribution and exceed global features in Low Block, Counter Attack, Def. Transition, and No Phase; global features lead in the other three phases.Figure 4 reports the per-class share of total Integrated Gradients importance attributed to pooled graph embeddings.Source: §5.1 · p. 9, Figure 4 · p. 10
  • Block phases share a dominant possession and ball-in-play state in 95.3%, 97.4%, and 98.1% of frames, so separating them requires spatial configuration.Under those states, the phase remains undetermined; the authors caution that Figure 4 measures relative reliance under its mean-embedding baseline, not each stream’s unique information.Source: §5.1 · p. 10
  • The t-SNE projection of learned graph embeddings shows classes appearing well separated, indicating phase-specific structure.The visualization projects 64-dimensional embeddings to two dimensions and displays a random 25% validation-set subsample.Source: §5.1 · p. 10, Figure 5 · p. 10

5.2 Relevance of Node Features

  • Horizontal position outweighs vertical position in every class; x-coordinate is the top feature in four classes, while team association leads in the other three.The x-coordinate reaches 40.1% attribution for High Block and 36.3% for Low Block; team association sums three one-hot columns, each contributing roughly 6–9%.Source: Table 4 · p. 11, §5.2 · p. 11
  • Angle to ball is the largest dynamic feature in all classes except No Phase, while distance to goal contributes little beyond x-coordinate.Speed and movement direction contribute moderately; the authors conclude that the model uses tactically plausible features consistent with domain knowledge of defensive phases.Source: §5.2 · p. 11

6 Discussion

The discussion finds that player relationships add predictive information beyond tabular features, with Delaunay graphs and SEAConv among the strongest design choices. It also highlights interpretability, practical reuse, and limitations in the evaluation and data scope.

6 Discussion

  • 4.6% lower macro F1 and 28.8% higher cross-entropy were achieved by the GNN-LSTM than the strongest baseline, XGBoost on standard features.The authors attribute the advantage to spatial relationships that tabular features do not capture.Source: §6 · p. 11
  • Delaunay triangulation outperformed alternative adjacency strategies, while complete graphs ranked last with 11.2% higher cross-entropy.Complete graphs also required 67% longer training time; the authors suggest removing distant-player edges reduces uninformative messages while multi-hop passing preserves information flow.Source: §6 · p. 11
  • SEAConv ranked first among tested message-passing operators, with 2.0% and 4.1% lower macro F1 than GCNConv and GINConv, respectively.Against GATConv, it tied in cross-entropy and had +0.005 macro F1; the authors leave its specific advantage and cross-dataset performance open.Source: §6 · p. 11
  • Graph embeddings received at least 36.1% of total Integrated Gradients attribution across classes and the majority share in four of seven.The global features alone cannot separate the three block phases, and the x-coordinate was the most important node feature next to team association; probabilities were well calibrated except for Low Block.Source: §6 · p. 12
  • The framework automates phase labeling and can be retrained on professional teams’ custom phase definitions.The graph representation may support pattern detection at individual, group, and team levels, while graph-based expected possession value models remain a potential downstream use.Source: §6 · p. 12
  • The comparison is limited by rule-generated labels, men’s football from a limited set of competitions, no independent test set, and no per-configuration hyperparameter tuning.The authors frame results as a controlled design-choice comparison under a fixed protocol, not optimized performance estimates.Source: §6 · p. 12

A Definition of Play Phases

  • The appendix formalizes six play-phase labels and specifies that a phase remains active until the ball goes dead.Pitch thirds are defined relative to the team whose phase is being described, with the final third adjacent to the opponent’s goal.Source: A Definition of Play Phases · p. 16

A.1 General Concepts

  • A player is stationary below the specified forward-or-backward movement threshold, while a team is organized when its unit-line score is below 250.The stationary threshold is 10 km/h; the organization score weights the rear and middle band widths four times each and the front width once.Source: §A.1 · p. 16

A.2 Build Up

Build Up is an in-possession phase defined by defender positioning, with opposition determined by specified pressing and proximity conditions.

A.2 Build Up

  • Build Up requires at least eight defenders behind the ball, regardless of the ball’s location by pitch third.The phase is in possession.Source: §A.2 · p. 16
  • Build Up is classified as opposed when the defending team is pressing and the closest defender is within 12 m, at least two are within 6 m, or one is within 4 m of the ball.Source: §A.2 · p. 16
  • The Build Up phase remains active until its criteria fail for 500 ms within a 2.5 s window.Source: §A.2 · p. 16

A.3 Low Block

Low Block is an out-of-possession phase requiring a stationary, organized defense and specified player counts across pitch areas.

A.3 Low Block

  • Low Block requires the defending team to be stationary and organized.The team must have been stationary for at least 1.6 s during the previous 2 s.Source: §A.3 · p. 16
  • At least four defenders must be behind the ball and organized or in the own box, at least five in the own third, and at least eight in the middle or own third.Source: §A.3 · p. 16
  • The phase remains active until its criteria fail for 400 ms within a 2 s window.Source: §A.3 · p. 16

A.4 Mid Block

Mid Block uses stationarity, defender organization and positioning, and exclusion of Low Block to define the phase.

A.4 Mid Block

  • Mid Block requires the team to have been stationary for at least 1.6 s during the previous 2 s.Source: §A.4 · p. 16
  • At least four defenders must be behind the ball and organized or in the own box, and Low Block must not apply.Source: §A.4 · p. 16
  • At least eight defenders must be in the middle or own third, and the phase remains active until its criteria fail for 400 ms within a 2 s window.Source: §A.4 · p. 16

A.5 High Block

High Block requires a stationary, organized defense meeting specified positional and possession-timing criteria; it remains active until those criteria lapse for a defined interval.

A.5 High Block

  • High Block requires at least 1.6 seconds of team stationarity in the previous 2 seconds, an organized defense, and specified defender positions and possession timing.The phase continues until its criteria fail for 400 ms within a 2-second window.Source: §A.5 · p. 16

A.6 Counter Attack

Counter Attack is an in-possession transition involving forward ball movement shortly after possession changes, with additional recovery, speed, and proximity criteria.

A.6 Counter Attack

  • Counter Attack requires forward ball movement within 3.5 seconds of a possession change, at least five recovering defenders, ball speed above 20 km/h, and no defender closer to the ball than an attacker.The possession-change timer does not increase during a counter attack; the phase remains active until its criteria fail for 400 ms within a 1.5-second window.Source: §A.6 · p. 16

A.7 Defensive Transition

This material defines defensive-transition labeling and describes baseline graph, sequence, topology, loss, and calibration methods, alongside figures and tables documenting evaluation materials.

A.7 Defensive Transition

  • Defensive transition is assigned to both teams after possession changes, with a defensive label for the team losing possession and criteria including an uncontrolled ball and no other in-possession phase.Ball control requires specified attacker and defender distances for 15 consecutive frames; the phase may persist through a simultaneous counter-press.Source: §A.7 · pp. 16–17

B Baseline GNN Layers

  • GCNConv and GINConv omit edge attributes, whereas GATv2 uses them to modulate dynamic attention weights rather than the messages themselves.GCNConv uses an unweighted adjacency in these experiments; GATv2 attention is conditional on the receiving node.Source: B Baseline GNN Layers · p. 17

C Long Short-Term Memory Networks

  • LSTM cells use input, forget, and output gates to update a cell state and produce an output, and are designed to mitigate vanishing gradients in sequential models.The forget gate controls retained state, while the input gate controls how much new candidate information is added.Source: C Long Short-Term Memory Networks · pp. 17–18

D Topological Features for the Baseline Model

  • Baseline topology features combine the teams’ convex-hull area ratio with persistent-homology entropies describing connected-component merging and loops in the 22-player formation.The area ratio is used as a measure of relative spatial dominance.Source: D Topological Features for the… · p. 18

E Cross-Entropy Loss

  • Cross-entropy is defined for binary and multiclass classification, with reported validation CE given as the mean per-observation loss.The multiclass formulation uses one-hot class indicators and predicted class probabilities.Source: E Cross-Entropy Loss · pp. 18–19

F Calibration Definitions

  • Calibration is assessed one class at a time by binning predictions and comparing observed positive fractions with mean predicted probabilities; weighted ECE is not comparable across classes.Because the ECE weights all observations, rare classes are dominated by many near-zero predictions.Source: F Calibration Definitions · p. 19

G Figures

  • Figure 7 shows residual validation macro-F1 distributions for adjacency and layer effects after removing the corresponding mean effects, with control settings marked.Higher values indicate better performance.Source: Figure 7 · p. 19

H Tables

  • The full-results table is sorted by Macro F1, while companion tables document paired comparisons, class-wise metrics, dataset composition, and attribution-analysis distributions.The pairwise tests compare adjacency strategies and message-passing layers; the class-wise metrics describe predictions from the Delaunay-adjacency, SEAConv model.Source: Table 5 · p. 20, Table 8 · p. 22, Table 9 · p. 22
Loading 2610.11571v1…