Using SHAP to analyze technical and tactical performance in table tennis based on hybrid Catboost-TPE with four-phase evaluation theory

Abstract This study adopted a novel algorithm, SHapley Additive exPlanation (SHAP), to analyze the table tennis matches based on a hybrid gradient boosting + categorical features-tree-structured parzen estimator (Catboost-TPE) with the four-phase evaluation theory. 110 singles’ matches (9536 rallies) were analyzed, and 59 elite male players’ winning rates from 2018 to 2022 were categorized into three levels (high, medium, low) by k-means cluster analysis. The results showed that Catboost-TPE has the best performance (MSE = 7.5e-05, MAE = 0.006, RMSE = 0.008, \({\text{R}}^{2}\)=0.99 and adjusted \({\text{R}}^{2}\)=0.989) among six hybrid machine learning algorithms. Using Catboost-TPE to calculate the SHAP value of each feature, the global interpretation and multiple local interpretations found that the performance of receive-attack and serve-attack phases have essential impacts on the winning probabilities in current matches. Besides, this study derived the mathematical equations for converting the scoring rate (SR), usage rate (UR) and technique effectiveness (TE) from the four-phase evaluation theory into the new three-phase evaluation theory to further deepen the theoretical and applied value of the four-phase evaluation theory used in this study. These results provided quantitative references to table tennis matches' characteristics and winning phases. These methods used in the study can be widely applied to other sports performance analyses, and the equations derived in this study are also instructive for relative sports.