Machine Learning Performance Analysis Not Transferable from Elite to University Football
A study from arXiv (2605.10796) investigates whether machine learning models trained on elite football data can be applied to university-level competition. Using event data from Europe's top five leagues and from National Tsing Hua University (NTHU), researchers trained Random Forest and Multilayer Perceptron models. Interpretability methods SHAP and CIS revealed that performance determinants learned from elite leagues do not transfer reliably to university football due to domain shift. The findings challenge the assumption that predictive models generalize across competition levels.
Key facts
- Study examines transferability of machine learning from elite to university football.
- Models trained on top five European leagues and applied to NTHU data.
- Random Forest and Multilayer Perceptron used with SHAP and CIS interpretability.
- Performance determinants are not structurally transferable across competition levels.
- Domain shift limits reliability of interpretations.
- Research published on arXiv with ID 2605.10796.
- Focus on interpretability, not just predictive accuracy.
- University data from National Tsing Hua University.
Entities
Institutions
- National Tsing Hua University
- arXiv
Locations
- Europe