Uma abordagem de aprendizado de máquina para identificação de tendências de sucesso de jogadores de basquete universitário americano para a liga profissional: conciliando predição e explicabilidade

Abstract

This study investigates the application of machine learning techniques on historical data from American college basketball (NCAA), with the aim of predicting which players have the highest potential to make it to the NBA. In recent years, data analysis in basketball has gained prominence, going beyond conventional statistics and incorporating data from sensors and cameras to identify patterns that can enhance player and team performance. However, although this growing complexity increases the predictive value, it also generates a dataset that may contain redundant, noisy, and irrelevant features, which can negatively impact predictive accuracy. To mitigate this problem and seek a solution that balances prediction and explainability of predictive models, this work proposes an approach composed of three main steps: (I) selection of the most relevant features to aid decision-making; (II) use of machine learning algorithms; (III) analysis of predictive results through the explainability of each model. In step (I), data selection was performed, and when combined, these features form a set of player attributes that directly or indirectly influence their hiring by NBA teams, considering the team’s current configuration. Established techniques from the literature, such as Wrapper, Filter, Embedding, as well as the Genetic Algorithm, were used to improve predictive accuracy and reduce the number of features. In step (II), the goal was to balance interpretability and predictive accuracy by employing transparent classification methods such as Decision Trees, Logistic Regression, and the Rule-based Algorithm (PRISM). The Support Vector Machine (SVM) was used as a reference for opaque models. In step (III), the explainability of each model was analyzed. This was done in two ways: through constructing the algorithms themselves, such as decision tree induction and rule-based algorithms, and through the SHAP explainability tool. For validation, the mentioned dataset was used, and the results indicated a positive impact of feature selection on the predictive models, with particular emphasis on the beneficial influence of the Genetic Algorithm in the selection phase. This approach contributed to identifying a minimal set of features and improving the prediction metrics of the classifiers. Specifically, combining the Genetic Algorithm with SVM in the fitness function during the selection phase produced a set of features that, when used in the Decision Tree (CART), achieved an accuracy of 80%. Finally, an analysis of the interpretability of the CART Decision Tree and PRISM models was carried out, highlighting the clarity provided by each: the first based on a decision tree structure and the second on a rule-based algorithm. Additionally, the SHAP tool was used to analyze the outputs generated by the machine learning algorithms, allowing for a clearer interpretation of the results, and thus aiding decision-making. Hopefully, these results will contribute to the efficient application of machine learning techniques in basketball, particularly in predicting future player behaviors based on explainable variables selected in a rigorous manner.