Explainable dual-task learning framework combining performance prediction and position classification for NBA player performance and position analysis
The majority of existing research on sports analytics currently uses machine learning methods which have become more common but they only analyze one specific task without providing ways to understand their model decisions. Our study proposes an explainable and integrated machine learning framework for predicting NBA player performance and classifying player positions. The proposed framework consists of three core components, a composite performance score (PERF\_SCORE) which Principal Component Analysis (PCA) uses to create an objective measurement system that removes all subjective weightings and a dual-task architecture which processes both regression and classification tasks and a SHAP (SHapley Additive exPlanations) driven iterative feature enhancement process. This paper compared four different ML algorithms such as Random Forest, XGBoost, Support Vector Machine and Linear Regression to see which one fits the model the best. The test results reveal that the combination of the various models in proposed approach brings better overall results in case of both training as well as testing phases, consistently across all models. In the XGBoost regression model, the R² value increased from 0.844 to 0.913, while the RMSE decreased from 2.95 to 2.57. The Random Forest model achieved 0.94 accuracy after starting from 0.86 in the classification task. The SHAP analysis method reveals global feature importance results while producing individual player results which make the model function as a clear decision support system. Our research presents a novel method which combines performance prediction with position classification into one system while using SHAP-based iterative enhancement as a core element. This system functions as an effective analytical resource for coaches, analysts, and sports decision-makers.
* [Engineering](/subjects/engineering)
* [Mathematics and computing](/subjects/mathematics-and-computing)
This work was supported by Recep Tayyip Erdogan University Development Foundation under grant number (02026006010390).
The authors declare no competing interests.