Improved Method for Evaluating Defensive Player Movements via Geometric Formation Features in Soccer

In recent years, geometric formation analysis using players' position information has advanced and demonstrated its usefulness in various team sports. However, its application to the evaluation of defensive player movements in soccer remains limited. Against this background, a previous study proposed a method for evaluating defensive player movements via geometric formation features. In this method, a statistical classification model that takes geometric formation features as input was used to predict the probability of goal concession given that an opponent's shot occurs within a certain time period. The defensive player movements were evaluated based on the prediction results provided by the model. However, for a defensive team, it is important to prevent not only goals but also shots themselves. Thus, the method had room for improvement in terms of the prediction of an opponent's shot occurrence. In this context, this study improves the method by applying an ordinal regression model that also predicts the probability of an opponent's shot occurrence. More specifically, the model is trained to predict the probabilities of ordered events, such as ''no shot,'' ''a shot off target,'' ''a shot on target without a goal,'' and ''a goal concession,'' based on an internal risk assessment score computed using the input geometric formation features. The improved method evaluates defensive player movements by comparing this risk assessment score with those of multiple computed counterfactual movements for each player. A validation experiment is also conducted to assess the effectiveness of the improved method.