SHAP과 LIME을 활용한 NBA 경기 승패 결정요인 분석

초록

This study applied interpretable machine learning to analyze the determinants of NBA game outcomes, with two research objectives: first, to verify whether a Stacking Ensemble outperforms individual models, and second, to simultaneously apply LIME and SHAP to interpret prediction rationale from multiple perspectives. Using NBA regular season data from the 2022-23 to 2024-25 seasons (3,690 games), six models—SVM, Decision Tree, Logistic Regression, Random Forest, XGBoost, and Stacking Ensemble—were compared. Results showed that the Stacking Ensemble achieved the best performance with a test accuracy of 85.1%, AUC of 93.3%, and Gap of 0.7%, outperforming all individual models. LIME analysis revealed that fg3m and dreb positively contributed to win predictions, while tov and pf were key factors in losses. SHAP analysis identified dreb, fg3m, fg3a, and pctFG2 as the most influential variables, and the interaction between dreb and fg3m statistically validated the modern basketball tactical pattern of defense–fast break–three-point shooting.

키워드