Enhancing Sports Performance with Big Data and Machine Learning-Powered Predictive Analytics

Abstract

In today’s football, the use of “Big Data” continues to rise in importance but there remains a substantial distance between the wealth of tactical information contained within fullmatch footage and what actionable, interpretable insights coaches can extract. Existing analytic work typically employs the type of discrete event statistics that do not provide information on the dynamic spatial-temporal development of team tactics across the entire full match period. What's more, the complex AI models of choice for such applications generally function as black boxes,making trust and applicability challenging. The latest new BIG-XAI model framework for the football performance analysis has been proposed with big data and XAI to resolve these shortcomings. The approach defines an end-to-end workflow starting from automatic analysis of fullmatch broadcast videos. The system is built upon a custom object detection model from fine-tuned YOLOv8 architecture. The DeepSORT (Deep Simple Online and Realtime Tracking) algorithm is used in conjunction to result stable storing of player trajectories. The pipeline also includes camera motion compensation through homography estimation to project player positions onto a top-down view of the pitch in order to estimate key spatial-temporal features, such as player speed, covered distance and team formations. More importantly, an XAI layer is embedded via saliency maps (Grad-CAM) and feature importance analysis (SHAP) allowing to transparently explainable descriptions of the models tactical insights in human language. Experiments show that the proposed framework is highly effective. A strong mAP@0. 5) of 0.677, and achieves excellent performance on tactically important classes: Player (AP: 0.990), Referee (AP: 0.995), Goalkeeper (AP: 0.871), and Goalpost (AP: 0.928). This robust detection allowed accurate player tracking,team determination and successful extraction of physical performance measures. The XAI module successfully converted complex model outputs to meaningful visual and textual explanations, effectively recognizing relevant events like possession turnover and putting in a broader context performance metrics such as ball possession for the coaching staff. In summary, this study represents an important step forward in providing full match tactical analysis through the development of a comprehensive interpretable AI framework which goes beyond historical events of play to deliver a dynamic, transparent and actionable profile of the evolution of tactics over the course of an entire competitive match. It connects machine learning sophistication and football coaching needs, establishing a new bar for trustable and acceptable AI in sports analytics.