하이퍼파라미터 최적화를 통한 머신러닝 모델 비교와 SHAP 분석을 활용한 KBL 경기 예측

초록

This study analyzed game data from the Korean Basketball League (KBL) during the 2021-2024 seasons to predict and evaluate game outcomes using various machine learning algorithms. Through hyperparameter optimization and 10-fold cross-validation, the XGBoost model demonstrated superior performance in game prediction compared to other algorithms, proving its high accuracy and generalization ability. The XGBoost model achieved an accuracy of 86.1%, a precision of 0.840, a recall of 0.896, and an F1 score of 0.867. According to SHAP analysis, defensive rebounds (DREB) and field goal percentage (FG%) were identified as the most important features. Turnovers (TO) negatively impacted game outcomes, while steals (STL) and offensive rebounds (OREB) were positive factors; however, their influence tended to diminish in the later stages of the game. These findings suggest that effective turnover management is essential for minimizing negative impacts during the latter part of the game. Furthermore, the study indicates that long-term improvements in three-point shooting percentage (3P%) and free throw percentage (FT%) can further enhance team performance.

키워드