This study adopts a new approach, SHapley Additive exPlanation (SHAP), to diagnose the table tennis matches based on a hybrid algorithm, namely Long Short-Term Memory–Back Propagation Neural Network (LSTM–BPNN). 100 male singles competitions (8535 rallies) from 2019 to 2022 are analyzed by a hybrid technical–tactical analysis theory, which hybridizes the double three-phase and four-phase evaluation theories. A k-means cluster analysis is conducted to classify 59 players’ winning rates into three levels (high, medium, and low). The results show that LSTM–BPNN has excellent performance (MSE = 0.000355, MAE = 0.014237, RMSE = 0.018853, and $${\mathrm{R}}^{2}$$
R
2
= 0.988311) compared with six typical artificial intelligence algorithms. Using LSTM–BPNN to calculate the SHAP value of each feature, the global results find that the receive-attack and serve-attack phases of the ending match have essential impacts on the mutual winning probabilities. Finally, case applications show that the SHAP can directly obtain each feature importance on one or more matches, which is more objective and reliable than the traditional simulation method. This research explores an innovative way to understand and analyze matches, and these results have implications for the performance analysis of table tennis and related racket sports.