Learning Spatial Formations in Football Using Graph Neural Networks and Contrastive Embeddings

This study examines the application of Graph Neural Networks (GNNs) in spatial analysis for professional football. Using high-frequency tracking data from Polish Ekstraklasa seasons, we constructed graph-based representations of player positions to capture dynamic interactions within teams. An overcomplete graph autoencoder with EdgeConv layers was trained using a contrastive learning objective, enabling the model to learn meaningful embeddings of team formations. Results demonstrate that the proposed approach effectively clusters similar match frames, highlights stable tactical structures, and identifies moments of rapid tactical change such as counterattacks. Compared to traditional spatial-temporal analyses, the GNN-based embeddings provide a more flexible, data-agnostic, and interpretable framework for tactical insights. This work contributes to the growing field of AI-driven sports analytics, with potential applications in real-time decision support, scouting, and training optimization.