Análisis y aplicación de Graph Neural Networks en el fútbol

Abstract

Abstract: This thesis proposes the use of Graph Neural Networks (GNN) for the analysis of passing networks in football, aiming to model player interactions and predict the receiver of the next pass. Using event data from Premier League matches, a complete pipeline was developed, including data preprocessing, intra-team graph construction, and the generation of temporal snapshots to preserve the dynamics of the game. The problem was formulated as a link prediction task, where different GNN architectures, specifically GCN, GraphSAGE, and GAT, were evaluated. The results, measured using the Hits@5 metric, demonstrate that the graph based approach is viable and capable of capturing meaningful structural patterns in the game. However, limitations were identified, particularly related to the absence of contextual variables and the size of the dataset. This work establishes a baseline for future research integrating spatial information and more advanced models.