Limited Transferability of Elite Football ML Models to University Competition
A new study published on arXiv investigates the transferability of interpretable machine learning models from elite European football leagues to university-level competition. While machine learning is widely used in sports performance analysis, most research assumes that performance determinants are consistent across different competition levels. This research challenges that assumption by training Random Forest and Multilayer Perceptron models on data from the top five European leagues and applying them to data from National Tsing Hua University. Using SHAP and Counterfactual Impact Score for interpretation, the study found that elite football exhibits stable performance determinant hierarchies. In contrast, university football showed significant reordering of key indicators, reduced explanation stability, and weaker structural agreement with elite domains. The findings suggest that interpretability robustness is domain-dependent. Instability in explanations under domain shift may indicate structural ambiguity in the target domain rather than just methodological limitations. This highlights the need for caution when applying elite-level analytical frameworks to lower-tier competitions, as the underlying dynamics of the game differ substantially between professional and university settings.
Wire timeline
Limited Transferability of Elite Football ML Models to University Competition
A new study published on arXiv investigates the transferability of interpretable machine learning models from elite European football leagues to university-level competition. While machine learning is widely used in sports performance analysis, most research assumes that performance determinants are consistent across different competition levels. This research challenges that assumption by training Random Forest and Multilayer Perceptron models on data from the top five European leagues and applying them to data from National Tsing Hua University. Using SHAP and Counterfactual Impact Score for interpretation, the study found that elite football exhibits stable performance determinant hierarchies. In contrast, university football showed significant reordering of key indicators, reduced explanation stability, and weaker structural agreement with elite domains. The findings suggest that interpretability robustness is domain-dependent. Instability in explanations under domain shift may indicate structural ambiguity in the target domain rather than just methodological limitations. This highlights the need for caution when applying elite-level analytical frameworks to lower-tier competitions, as the underlying dynamics of the game differ substantially between professional and university settings.
cs.AI updates on arXiv.org