Beyond the Table: A Federated Learning CNN’s Framework for Classifying Serve Shots in Table Tennis

The approach categorized serve shot motions into five distinct classes, enabling detailed performance analysis for players and coaches to rely on the Federated Averaging (FedAvg) method to achieve the aim of our endeavor. This allowed us to aggregate local data learning into a unified global model. This study indicates that the data, which is international and the utility of which is further improved by the inputs of clients from different locations, has achieved significant levels of recall, precision, and accuracy, respectively. Client bj_5 achieved outstanding levels of precision and recall, with both metrics reaching 88.86% and 89.07%, respectively. The precision-recall score reached 0.89, while the accuracy was 0.96. In contrast, bj_2 exhibited a minor delay, achieving precision and recall rates of 76.96% and 77.21%, respectively. In that respect, the system got an F1-Score of 77.01% and an accuracy of 91%. Additional macro, micro, and weighted averages examination highlighted the model's reliable performance across different serve shot motions. The highest value of the macro average (i.e., 77.06%) was observed in model bj-2, while the lowest was recorded in model bj-5 (i.e., only 83.35%). Micro averages of the model showed a more comprehensive range of model accuracy, which was only 76.99% in the model bj-2, and only the model's robustness was confirmed by observing weighted averages ranging from 76.94% (bj_2) to 88.94% (bj_5). Results reveal the intervention as an opportunity for the branch of sports statistical analysis to create accurate predictive models while at the same time safeguarding athlete privacy. Our work not only enhances the set of methods used for analyzing athlete performance but also offers a scalable approach that may be used in other fields, including processing and analyzing sensitive data.