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

arXiv:2610.11571 · stat.AP, stat.ML · Submitted 2026-10-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Automated Detection of Match Phases in Football from Spatio-Temporal Tracking Data Using Graph Neural Networks".

Jane: The gist: 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,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: So we're looking at this paper called "Automated Detection of Match Phases in Football from Spatio-Temporal Tracking Data Using Graph Neural Networks." It’s essentially taking all that second-by-second positional tracking data from football matches and trying to figure out what phase the game is in.

Jane: That sounds pretty complex, Tom. So, instead of just looking at where players are on the field at one time, this paper is using a graph neural network to model how all those players interact over time and classify the match phase across seven different categories.

Lu: Yeah, it’s about modeling these interactive movements because that’s what makes football tactically meaningful; you can't just look at individual positions in isolation >

Meng: So, what exactly is the core idea here? How are they going to handle all those players and the ball data to get a phase classification on every single second?

Tom: Well, the paper proposes combining graph neural networks with a sequential model. They build this directed graph where each node is a player or the ball, and the edges describe their connections >

Jane: And they’re not just using any connection method; they specifically construct the adjacency matrix using Delaunay triangulation because that mirrors how football moves spatially on the pitch >

Lalam: That Delaunay construction connects nodes into triangles such that no other node lies inside the circumcircle of any triangle, which helps capture passing lanes structure better than some alternatives >

Tom: Right, so they compare this to other ways of building the graph, like using k-nearest neighbors or a complete graph, and they found that Delaunay triangulation performs best for adjacency construction >

Jane: And the model uses a message-passing layer called Spatial Edge-Augmented Convolution, or SEAConv; it’s designed to inject edge attributes directly into messages to capture spatial dependencies while limiting uninformative messages from redundant edges >

Meng: So they’re not just aggregating features into team-wide summaries; they are using player-level features directly, which means the system preserves the full spatial structure of all players instead of aggregating them >

Title and authors: Tom: Exactly. They also use this sequence of graph representations and then feed that through a unidirectional LSTM to classify the phase at each timestep, which lets them capture temporal dependencies among football players >

Lu: The data itself is quite detailed; they are using positional data at twenty-five Hertz for all twenty-two players and the ball, along with kinematic features like speed and acceleration >

Jane: And that means every second of the match has very specific spatial and movement details available to feed into this GNN architecture >

Tom: So, they tested this system on a validation set using accuracy, balanced accuracy, and macro F1 score for seven different match phases >

Meng: The results show that their GNN approach outperforms every baseline they compared against on every metric; even XGBoost on the lagged standard features trails by only two point four percent in accuracy >

Jane: And when you look at the breakdown, it shows a three point seven percent drop in balanced accuracy and a four point six percent drop in macro F1 score compared to that baseline >

Tom: The paper also looks closely at the design choices, running eighty different tests on adjacency constructions and message-passing layers to see what works best >

Lu: And they found that SEAConv achieves the best performance on every metric except for the CE loss, where it’s only marginally different from GATConv >

Jane: The interpretation section uses Integrated Gradients to show which features matter most; they found that graph embeddings receive at least thirty-six point one percent of the overall attribution across all seven classes >

Tom: And something really telling is that for phases like ‘High Block’ and ‘Low Block’, the x-coordinate was the most important input feature, next to team association >

Meng: So, it suggests that horizontal positioning directly distinguishes defensive phases based on how deep the defense is set on the pitch >

Lu: That makes sense because spatial relationships between players carry predictive information for match phase classification that simple tabular feature representations don't capture >

Title and authors: Jane: It’s a big deal because this framework automates phase labeling, which used to be something done manually by human analysts in every single match >

Tom: But the paper also flags some limitations; they see performance drops for transition classes like ‘Counter Attack’ and ‘Defensive Transition’ where the F1 scores are low, specifically zero point five four and zero point six six respectively >

Meng: They noted that this happens because those phases are typically very short, and the labels they used were generated by a rule-based system rather than being manually annotated >

Jane: It’s a fair point; the method struggles with those quick, transient moments where defining the phase is tricky based on how it was labeled >

Tom: Overall, this paper calls "Automated Detection of Match Phases in Football from Spatio-Temporal Tracking Data Using Graph Neural Networks" and it suggests that domain-informed graph construction and edge-aware message passing might benefit other tactical tasks using tracking data >

Lu: I think the ability to capture those spatial relationships is the key insight here, moving beyond just looking at isolated points on a field >

Jane: So, to wrap up, this framework shows how combining GNNs with a sequential model can effectively classify match phases on a second-by-second basis across seven classes >

Tom: It’s a solid comparison of design choices under fixed training protocol, but those transition phases are where the current method hits its limits >

Meng: I see this having practical relevance for things like real-time tactical analysis or even automated scouting reports, though getting those short phases right will be hard >

Lalam: From an LLM perspective, this capability means we can build systems that understand complex spatial contexts in dynamic environments far more accurately than just reading static text descriptions >

Lu: I think the future work should focus on making those transition class predictions more robust, maybe by incorporating different types of temporal context into the LSTM processing >

Jane: It’s a fascinating look at how graph structures can unlock tactical meaning in complex physical data like football tracking, and it sets a good benchmark for other sports analytics research >

The paper's summary: Tom: So, this paper takes all that second-by-second positional tracking data from football matches and uses a graph neural network to classify which phase of play is happening across seven different categories.

Jane: It’s basically building a map of how the players are interacting on the field over time so an AI can tell you what's going on at any given moment.

Tom: Yeah, the authors compare their system against all these standard ways to model player movement and they found that their combination of graph structure and sequence modeling actually beats every single one of those baselines.

Jane: It’s not just a little better; they showed the GNN approach outperforms XGBoost on every metric, even though XGBoost is usually considered pretty strong for tabular data.

Tom: And it gets really interesting when you look at *why* it works because they use integrated gradients to show that the spatial relationships—the graph embeddings—are actually contributing a big chunk of the prediction, over thirty-six percent across all those phases.

Jane: That means it’s not just about what a player is doing in isolation; it's about how they are positioned relative to everyone else on the pitch.

Tom: Right, and they pinpoint something specific for certain phases like 'High Block' and 'Low Block': the x-coordinate was the most important input feature, which makes total sense because horizontal positioning tells you how deep the defense is set.

Jane: So what this changes for us is that we can automate labeling in a way that actually respects the physical layout of a game, not just some simple summary of statistics.

Tom: Exactly, but there are some limitations they admit too; they said performance dips for those really quick phases like 'Counter Attack' and 'Defensive Transition' because those moments are so short, and the labels used were based on a rule-based system rather than being manually annotated.

Jane: That’s a fair caveat; the method struggles when the event itself is too transient to be captured properly by their labeling rules.

Tom: So they’re suggesting that while this framework automates phase labeling for things like tactical analysis, we still need to figure out how to make those short transition phases more robust in the future.

Jane: It really opens up possibilities for real-time scouting or even training simulations where you can instantly label what's happening on the screen.

Tom: And this whole thing is just a starting point; they think that using domain-informed graph construction and edge-aware message passing could actually help other tactical tasks that deal with tracking data in the future.

The paper's improvements: Tom: So, we’re looking at how they suggest improving this model because they noticed some weaknesses in the original design for certain phases, right?

Jane: Yeah, they’ve put forward a few ideas on how to make it better than just using the basic setup.

Tom: One big idea is to use a Graph Neural Network with this custom Spatial Edge-Augmented Convolution layer, or SEAConv; that injects edge attributes directly into messages and captures spatial dependencies while cutting out those redundant messages from edges that don't matter.

Jane: That sounds like they’re making the connections between players smarter by adding extra context to each line of communication in the graph.

Tom: Right, and they claim this helps capture patterns based on pairwise distance, relative angle, and team membership better than the original GINConv setup, which leads to a performance gain of plus zero point zero three two five in macro F1 score.

Jane: So they’re saying that by making the graph connections more specific to what actually matters in football tactics, you get a measurable lift in accuracy across the board.

Tom: They also propose using player-level features directly, instead of just summarizing everything into team-wide features; this keeps the full spatial structure of all players intact, which lets them identify phases based on actual spatial arrangements.

Jane: That means they’re focusing on how players are spread out across the pitch at that exact second rather than just looking at a general team statistic.

Tom: And they link that directly to the interpretation stuff we talked about earlier; Integrated Gradients showed that for phases like 'Low Block' and 'High Block', the x-coordinate is still super important, which confirms this focus on horizontal positioning is key.

Jane: It reinforces that the model isn’t just guessing; it’s learning to look at the physical layout of a defense to make its decision.

Tom: On the modeling side, they are using a unidirectional LSTM layer to process those graph representations sequentially, which is what lets them classify the match phase on a second-by-second basis by capturing temporal dependencies among the players.

Jane: So they’re combining spatial structure with time sequence so it can see both where everyone is and how that situation evolves over time.

Tom: And looking at the limitations, they pointed out that those transition phases like 'Counter Attack' and 'Defensive Transition' still have low F1 scores because those events are just too short to be reliably labeled by their current rule-based system.

Jane: It’s a practical problem; you can’t have a good classification method if it fails on the very quick, high-action moments that define transitions.

Tom: The implication here is that while the foundation of spatial modeling is strong, the next step in improving this kind of system has to be handling those very brief, high-velocity events with more flexible temporal context in the model.

Conclusion: Tom: So, to wrap up, this paper on "Automated Detection of Match Phases in Football from Spatio-Temporal Tracking Data Using Graph Neural Networks" shows that combining spatial graph structures with sequential modeling is a solid way to get accurate phase classification across seven different categories.

Jane: It’s really about proving that the physical layout of the game matters more than just looking at isolated player statistics when you’re trying to understand what's happening on the field.

Tom: Yeah, and they did this by showing that their GNN approach beats every other baseline they tested, even though it's a lot more complex than simpler methods.

Jane: The big picture is that this moves us closer to systems that can actually interpret complex physical scenarios in real time rather than just spitting out numbers.

Tom: They did flag those transition phases like 'Counter Attack' because those short events are tricky for the current rule-based system to get right, which means future work needs to focus on making the model more flexible for those rapid shifts.

Jane: It’s a fair critique; real-world data is messy and quick, and if we want this tool to be useful in scouting or coaching, it has to handle those fast moments better than the authors did now.

Tom: This work really establishes a good benchmark for using graph neural networks on dynamic sports tracking data, but it’s not the end of the road yet.

Jane: Exactly; once we nail how to handle those quick transitions, this framework could be incredibly useful for any sport where spatial movement is key.

Tom: Lu thinks that using these types of spatial-temporal models could be really powerful for understanding complex human interaction patterns in any dynamic environment, not just football.

Jane: I agree; it opens up avenues for how we build AI that understands context, not just single data points.

Tom: Meng sees the practical side here; if this can reliably label phases, it means a coach or analyst could get instant feedback on defensive setups based on where players are positioned, which is a tangible application.

Jane: That’s true; it translates complex tracking data into actionable intelligence for people who actually need that insight.

Tom: Lalam feels this kind of spatial reasoning ability is crucial for developing large language models that can truly grasp the physical world and help us build a more nuanced understanding of human activity in any context.

Tom: So, "Automated Detection of Match Phases in Football from Spatio-Temporal Tracking Data Using Graph Neural Networks" gives us a strong proof that graph-based AI is effective for spatio-temporal sports analysis.

Jane: It shows that the way players are connected on the field tells a much better story than just looking at their individual locations.

Tom: And while it has limitations with very short phases, it sets up a clear path for refining those models to handle complex, fast action sequences in the future.

Lu: I think this is foundational work because modeling these spatial dependencies correctly is something we can apply to so many other dynamic systems beyond sports.

Meng: From an engineering standpoint, the comparison against those standard baselines shows that even with complexity, we found a good trade-off between performance and computational cost for the Delaunay construction.

Lalam: For me, this kind of deep spatial understanding helps our underlying models develop a richer cultural context about how different elements—players in this case—relate to each other across time.

Vincent Renner, Nils Koster, Pascal Bauer, Melanie Schienle

Institute of Statistics, Karlsruhe Institute of Technology · Karlsruher Sport-Club · Broad Institute of MIT & Harvard · German Football Association (DFB) · Chair for Sports Analytics, Saarland University

stat.AP, stat.ML

Submitted: 2026-10-08

Updated: 2026-10-08

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 79/100

The gist: The gist: 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,

Key concepts

Graph Representation (Gt)
The football field is modeled as a directed graph where each node represents a player or the ball. Edges connect players based on their spatial proximity, typically constructed using Delaunay triangulation to capture the natural spatial structure of the pitch.
Spatial Edge-Augmented Convolution (SEAConv)
This is a specialized message-passing layer for GNNs that enhances standard convolution by incorporating information from the edges connecting nodes. It improves performance by augmenting both the messages passed between nodes and their attention coefficients, allowing the model to better understand local spatial interactions.
Integrated Gradients (IG)
This is a technique used to interpret GNN predictions by calculating feature importance. It shows which input features—like graph embeddings or specific coordinates—most influenced a particular classification. This helps explain *why* the model made a specific prediction about the match phase.
Sequential Model (LSTM)
After the GNN processes player data into sequences, a unidirectional Long Short-Term Memory (LSTM) network analyzes these sequences over time. This allows the system to capture temporal dependencies and classify match phases based on how player positions evolve second by second.

Terminology

Summary

The gist: 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, outperforming all aggregated-feature baselines.

Data and Graph Representations

The data set comprises positional data at 25 Hz for all 22 players and the ball, described by spatial and kinematic features such as speed and acceleration The paper constructs a directed graph Gt = (V, E) where V consists of all players and the ball, with k = V = 23 Each node vi carries an m-dimensional attribute vector xi ∈ R m comprising spatial coordinates, movement descriptors, and indicators for the node type relative to the focal team Connectivity is encoded by an adjacency matrix A ∈ 0, 1 k×k with ai,j = 1 if edge (vi, vj) ∈ E The authors propose constructing the adjacency matrix through Delaunay triangulation (Delaunay, 1934), motivated by the spatial structure of football This construction is compared against three general-purpose strategies: k-nearest neighbors (k-NN) graph connects each node to its kNN spatially closest nodes, using kNN ∈ 3, 6 in our evaluation setting, the radius graph connects all node pairs whose Euclidean distance falls below a threshold r, equivalent to the threshold-based sparsification used by Raabe et al. (2023), and the complete graph connects every node to every other node, making no structural assumptions but incurring the highest computational cost.

Model Design

The proposed GNN uses a generic message-passing layer update rule where x′i = γΘxi, Mj∈N(i)ϕΘ(xi, xj, ej,i), where L is a differentiable, permutation-invariant aggregation The authors specifically propose the Spatial Edge-Augmented Convolution (SEAConv) layer to address two limitations of existing operators SEAConv adapts the edge-augmented graph transformer layer of Shi et al. (2021) into a message-passing layer in which edge embeddings augment both the attention coefficients and the messages To resolve fine-grained positional differences, spatial coordinates xposi ∈ R 2 are encoded via random Fourier features ψ(·) ∈ R 2m′ The resulting sequence of graph representations Hseq = (Z1∥ϕu(u1), Z2∥ϕu(u2),..., ZT ∥ϕu(uT)) is then processed by a unidirectional LSTM This sequence is normalized and transformed by a learned projection before being passed to the recurrent model.

Empirical Results and Comparison

The GNN models are evaluated on the validation set using accuracy, balanced accuracy BAcc, and macro F1 score F1macro The GNN outperforms every baseline on every metric The strongest baseline is XGBoost on the lagged standard features, which trails the GNN by only 2.4% in accuracy but by 3.7% in balanced accuracy, 4.6% in macro F1 and 28.8% in CE. The authors systematically compare adjacency constructions and message-passing layers across 80 runs to isolate the effect of each design choice Delaunay triangulation performs best for adjacency construction, followed by k-NN (kNN = 3) graph SEAConv achieves the best performance on every metric except the CE loss, where the difference to the GATConv is marginal.

Model Interpretability

The model’s predictions are made interpretable through Integrated Gradients (IG; Sundararajan et al., 2017) to analyze the importance of graph’s features The analysis shows that graph embeddings receive at least 36.1% of the overall attribution across all classes, while global features take a larger share in ‘Build Up’, ‘Mid Block’, and ‘High Block’. The x-coordinate was the most important input feature next to the team association, which makes tactical sense as horizontal positioning directly distinguishes defensive phases at different pitch depths.

Discussion and Limitations

The central finding is that spatial relationships between players carry predictive information for match phase classification that tabular feature representations do not capture The proposed framework offers practical relevance as it automates phase labeling, a task currently performed manually Limitations include performance drops for transition classes ‘Counter Attack’ (F1 = 0.54) and ‘Defensive Transition’ (F1 = 0.66) because these phases are typically very short, and the labels are generated by a rule-based system rather than manual annotation. The results should be read as a controlled comparison of design choices under a fixed training protocol rather than as optimized performance estimates. The findings suggest that domain-informed graph construction and edge-aware message passing may benefit other tactical tasks operating on tracking data.

--- Page 1 ---

The gist: 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, outperforming all aggregated-feature baselines.

Model Interpretability

The model’s predictions are made interpretable through Integrated Gradients (IG; Sundararajan et al., 2017) to analyze the importance of graph’s features The analysis shows that graph embeddings receive at least 36.

Improvements for AI systems

  1. Use a Graph Neural Network (GNN) with a custom Spatial Edge-Augmented Convolution (SEAConv) layer to model player interactions, which inject[s] edge attributes directly into messages and captur[es] spatial dependencies while limiting uninformative messages from redundant edges. This allows the system to distinguish between tactical patterns based on pairwise distance, relative angle, and team membership, leading to a performance gain of +0.0325 in Macro F1 over GINConv.

  2. Implement a model that uses player-level features directly, preserving the full spatial structure of all players instead of aggregating into team-wide features, which allows for the identification of phases based on spatial player arrangements. This enables the system to leverage input from horizontal positioning and other spatial configurations, as indicated by Integrated Gradients showing that these features are critical for separating phases like ‘Low Block’ and ‘High Block’.

  3. Incorporate a sequential model, specifically a unidirectional LSTM, to process the sequence of graph representations, as this architecture is used to produce the final classification P(yt = c G≤t) at each timestep. This allows the system to capture temporal dependencies among football players and classify match phases on a second-by-second basis.

  4. Utilize domain-informed graph construction by using Delaunay triangulation for adjacency, as this strategy is shown to perform best, with the paper stating that it encode[s] this property [of triangles] by connecting nodes into triangles such that no other node lies inside the circumcircle of any triangle. This improves the model's ability to capture passing lanes structure compared to alternatives like complete graphs.

  5. Employ Integrated Gradients (IG) to provide model interpretability, specifically analyzing the importance of the graph’s features, as well as the overall contribution of individual spatial features compared to graph-wide possession-based information. This analysis reveals that for 'High Block' and 'Low Block' phases, the x-coordinate was the most important input feature, providing domain-specific insight into how horizontal positioning distinguishes defensive phases.

Abstract

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.

Sources

Related papers