Graph Neural Networks for Football Match Analysis and TacticalStyle Clustering

Sammanfattning

Football analytics has recently shifted toward graph-based methodologies to handle thecomplex, relational data generated by player interactions. While traditional match analysisrelies on aggregate statistics or standard machine learning, these methods often treat matchevents as isolated snapshots. This approach misses the broader, structural logic of teamplay. To address this, we represent matches as passing networks, defining players as nodesand their passing exchanges as edges We argue that this graph representation is not justa technical alternative, but a necessary shift for capturing the multi-agent dynamics of thesport.Our research follows a clear logic: before predicting match outcomes, we must first assess individual performance. We move away from conventional metrics by using a graphbased framework to calculate node centrality and interaction influence. This allows us toquantify a player’s role in orchestrating attacks or maintaining defensive shape, providinga granular view that summary statistics cannot capture.For matchoutcome prediction, wetestbothGraphConvolutionalNetworksandGraphAttention Networks to determine which better captures structural patterns . Our findingsshow that GAT architectures which use dynamic attention to weight specific interactionsconsistently outperform isotropic GCN models in prediction stability and accuracy.Finally, we use a Graph Autoencoder (GAT-GAE) to identify tactical styles. Instead ofrelying on rigid, pre-defined metrics, this approach uncovers structural patterns like hubconcentration and role differentiation, which often remain hidden under standard statistical analysis. Our results demonstrate that these graph-based embeddings produce morecohesive clusters, confirming that such models are essential for interpreting the underlyingtactical logic of modern football