Prediction of the national basketball team candidate player roster using machine learning algorithms

The present study aims to provide a solution to the problem of selecting players who will exhibit high performance in the national team according to season or match statistical attributes among players in the Turkish Women's Basketball Super League. This study is important to evaluate all players according to their statistical attributes while determining the national team candidate rosters. To this end, a range of classification algorithms – namely Naïve Bayes, Random Forest, Support Vector Machines, Extreme Gradient Boosting and Decision Trees – were applied to a data set comprising basketball statistics. The model with the highest accuracy and F1 Score was then created. The study revealed an imbalance between the number of Turkish women basketball players in the Women's Basketball League and the number of players who are candidates for the national team. While there are many players, the candidate roster represents a narrower group. The sampling algorithms applied in this study are designed to address this imbalance and insufficiency of data. In the study, firstly, a Principal Component Analysis was performed on the attributes consisting of statistical data. This analysis was performed with four different variance values, and the important ones of the attributes were preserved. Subsequently, before proceeding to the classification stage, the challenges posed by the small and imbalanced data set were re-evaluated using a range of oversampling methods, namely Synthetic Minority Oversampling, Random Over Sampling, Borderline SMOTE, Random Under Sampling and Random Under Sampling 5% + SMOTE. The modelling phase was then undertaken by implementing classification algorithms. The modelling process yielded the conclusion that the XGBoost algorithm functions with 86% accuracy and an F1 Score when employing the random oversampling method. The reliability of the models was validated through 10-fold cross-validation. It is hypothesized that the findings of this thesis study will contribute to studies based on player selection, such as determining the national team candidate roster. In future studies, it is anticipated that these models will be developed as software and will contribute to the evaluation of the performance of players while forming the rosters of both national teams and teams. The evaluation of the individual and interconnected situations of the players is also expected to contribute to the country's sporting success, facilitated by a performance-oriented modelling approach that employs numerical values.