An AI framework for counterattack detection and decision-making evaluation in football

This study proposes a performance analysis framework for evaluating counterattack decisions in football by utilizing deep learning techniques. A dataset of 101,710 frames was selected based on specific algorithmic rules from synchronized StatsBomb event data and tracking data of 580 Premier League matches. Subsequently, a comprehensive approach integrating Transformer and Graph Neural Networks was employed to model and predict match event decisions based on prior match events and tracking data. The results demonstrated that there are approximately 10 counterattacks per match. Each counterattack sequence is typically comprised of 5 events, and 2 of these sequences usually result in shot attempts (successful counterattack). Half of the counterattacks were initiated by defenders, with a success rate of 7.49%. Moreover, no significant monotonic correlation was found between the number of counterattacks per game and the number of goals scored. After modeling and inspecting Permutation Feature Importance, it was shown that player positions, distance advanced, and the relative angle to the carrier are the features with the greatest impact on model performance. Additionally, illustrative examples from the models were provided to aid in the effective analysis of player decision-making during such tactics. The findings of this study can offer strategic insights for coaching and gameplay improvement.