Analysis of multi-modal fusion momentum applications within the interpretable framework of SHAP

In tennis matches, we often witness surprising turnarounds in seemingly dominant players, which are usually attributed to the "momentum". This process of accumulating momentum over time and reaching a crucial point is known as "momentum". This "momentum" is difficult to quantify in matches, so our team has developed an interpretable "momentum" quantification model to provide valuable information for match teams and coaches, helping them achieve outstanding results. We identify the player's match turnaround points based on the change in momentum values, and use XGB, SVM, GNB, RF, MLP, and other 10 machine learning and deep learning algorithms to build a classification model with turnaround points as training labels. AUC, accuracy, recall, and F1-score are used as evaluation metrics, with K-fold cross-validation as the validation method. Eventually, XGB and RF are selected for multi-model weighted fusion, showing improvements in AUC, accuracy, recall, and F1-score compared to the other eight algorithms, with values of 0.98, 0.93, 0.98, and 0.94, respectively. Moreover, our model demonstrates good stability and robustness across different match datasets.To analyze the player's match momentum changes and enhance the algorithm's interpretability, we propose an interpretable algorithm framework based on SHAP. SHAP is a game-theoretic interpretable method that provides an explanation value for each feature and sample, indicating its relative contribution to the final prediction. This helps us understand how the model makes decisions based on different features, enabling coaches and teams to analyze match trends and provide valuable information more effectively.