한국프로야구 골든글러브 수상자 예측모델 비교

Abstract

Baseball has a wide variety of statistics compared to other sports, and these records are perfectly preserved, allowing for a detailed review of game contents and outcomes based on the statistics of each game. In baseball, there is an award called the “Golden Glove” given to the best player in each position. The selection of the Golden Glove winners is based on votes from media personnel such as reporters, broadcast PDs, and commentators, making it difficult to establish clear criteria and important variables for each position. Therefore, the aim of this study is to design a predictive model for the Golden Glove winners based on the statistics of candidates and winners from the Korean professional baseball league between 2003 and 2022, using logistic regression analysis and machine learning techniques. The study also compares and analyzes the performance of the designed predictive models to identify the most suitable model for predicting the Golden Glove winners. The models include logistic regression, support vector machines, random forests, and XGBoost, with hyperparameters specified for each model. Z-scores, which standardize values, and min-max normalization, which scales values, were used to present the models in two different forms. After optimizing variables, the performance of each model was evaluated. First, in designing the Korean professional baseball Golden Glove prediction model, the logistic regression model used L1, L2, and elasticnet as kernels, the support vector machine model used rbf and poly as kernels, and no significant variables were identified in the nonlinear models. The random forest model used gini and entropy as criteria, and the XGBoost model used exact, approx, and hist as criteria. To increase the F1 score, variables were removed one by one, and the selected variables for each position differed between models. Second, the comparison of the machine learning prediction models revealed that although there were differences in performance across positions, the support vector machine model 2 and XGBoost model showed high prediction accuracy and F1 scores, while the logistic regression and random forest models showed relatively lower accuracy and F1 scores. Overall, the models that were standardized using min-max (model 2) demonstrated better prediction abilities than those standardized using z-scoring (model 1). Therefore, to predict the Golden Glove winners, it is advisable to use the XGBoost model 2 with min-max normalization, and it is believed that including defensive statistics in future models will yield even better results.