Extended xThreat: an explainable quality assessment method for actions in football using game context

Engineering of Sport 15 - Proceedings from the 15th International Conference on the Engineering of Sport (ISEA 2024)

In the last decade, systematic collection of data is increasingly being used in the world of football which enables the use of mathematics to improve the performance of individual players and teams. Mathematical models allow for quality assessment for in-game actions by estimating the probability that a goal is scored. These models must be explainable, as they should be reducible to a few clear instructions that a coach can give to a player. Therefore, there is a need for explainable quantification methods for actions in football. Most models for action quality assessment can be described using either the Valuing Actions by Estimating Probabilities (VAEP) or Expected Threat (xThreat) frameworks. VAEP models typically use machine learning methods, which comes at the expense of explainability. On the other hand, the xThreat framework is explainable but it uses only the position of the player in possession of the ball as input. This means football data analysts using state-of-the-art techniques face a trade-off between an explainable but simplified xThreat model or a VAEP model that takes contextual variables into account. The main goal of this paper is to create an extended xThreat model that can include game context, thus aiming for explainability while taking into account contextual variables.