Publikation
Explainable Multi-Modal and Local Approaches to Modelling Injuries in Sports Data
Dan Hudson; Ruud JR Den Hartigh; L Rens A Meerhoff; Martin Atzmueller
In: 2023 IEEE International Conference on Data Mining Workshops (ICDMW). IEEE International Conference on Data Mining Workshops (ICDMW-2023), December 4, Shanghai, China, Pages 949-957, IEEE, 2023.
Zusammenfassung
To tackle the difficult task of predicting injuries, initial attempts have been made to apply machine learning methods to complex, multivariate sports data. However, no machine learning model has emerged which has good performance and generalisability to multiple datasets. Focusing on explainable machine learning, this paper discusses possible limitations in common approaches to injury prediction, and investigates methods that might be able to overcome these limitations. In particular, we first examine the value that can be obtained by including daily records of psychological variables to provide a multi-modal perspective on athletes. Secondly, we consider whether models that focus on learning from local rather than global patterns in the data are better suited to modelling injuries. We explain how the diversity of injuries in terms of severity, duration, type (chronic versus acute) and context may make local models preferable. The local modelling approaches we consider include subgroup discovery, anomaly detection and casebased classification. Since the proposed approaches threaten to increase the model complexity, we take great care to ensure that explainability is maintained, and describe the ways in which each of our methods is explainable. This leads to a set of explainable, local models. To provide an empirical basis for the investigations, we perform a series of computational studies using data obtained by daily recording of professional association football players in the Netherlands over a period of two seasons. These studies aim to expand the currently limited literature on applying explainable ML methods to injury prediction. Our results indicate that predicting injuries is very difficult in general, while the focus of this paper is mainly on providing insights on explainability and interpretability. Therefore, alongside results detailing the performance of various machine learning models, we also include multiple examples of explanations that can be extracted using our proposed methods, which may support actionability.