Predicting Emotional Contagion in Football: A GNN-LSTM Approach

In the context of sport, football is one of the more complex team sports where interactions and emotions spread dynamically across players. Understanding and predicting these interactions can provide valuable insights for performance analysis; however, traditional meth ods struggle to capture both the relational and temporal nature of the game. This thesis aimed to develop and evaluate a model capable of predicting players’ interactions and emotions by combining graph based and sequence-based learning methods. A GNN-LSTM model was proposed to capture spatial dependencies between players and temporal dynamics across match sequences. The structure was first built on semi-synthetic data and then tested and evaluated on the real datasets. Additional experiments compared different loss functions to assess their impact on imbalanced data in an attempt to improve the performance. The proposed GNN LSTM model achieved a weighted average of approximately 85%, outperforming the baseline models of 72% and GNN alone of 80%. The greatest improvements were observed in the minority classes, which were previously difficult to predict. Loss function experiments showed that standard Cross-Entropy remained the most effective, with differences between losses below 2%. This work demonstrates that combining GNNs and LSTMs provides a powerful framework for spatio-temporal predictions in football, offering both improved classification and insights into individual player influence through graph analysis.