Predicting Football Player Cooperation Using Transformer Models: A Comparative Study with Traditional Machine Learning Approaches

Accurately predicting whether football players can work well together on the field is essential for optimizing team strategies, improving match outcomes, and making decisions in player transfers. This is not an easy task, player interactions are complex, constantly shifting based on roles, tactics, the opposing teams strategies, and the changing dynamics within the squad over time. However, in football, there are many data and statistics; there must be a hidden relationship that lies below them; this exactly falls within the scope of machine learning. This paper introduces a Transformer-based model designed to capture these temporal interactions between players. Using the self-attention mechanism, the model is able to focus on key moments of cooperation. Our approach combines player performance data with custom cooperation metrics to predict which players are likely to work well together in upcoming seasons. Extensive experiments were conducted, comparing the Transformer to traditional methods such as Support Vector Machine (SVM), Random Forest, and Logistic Regression, using real-world football datasets that include player statistics, season ratings, and positional information. While the Transformer achieved a validation accuracy of 75.31%, it was slightly outperformed by the Random Forest model, which reached an accuracy of 77.78%. Nevertheless, the Transformer model demonstrates the ability to model temporal interactions effectively, indicating potential for further improvements.