Sports Analytics Using Data Mining, Probabilistic Reasoning, and Deep Learning
Abstract:
This thesis presents a multi-disciplinary research that integrates data mining, probabilistic reasoning, and deep learning to address challenges in sports analytics. The primary goal is to provide in-depth strategy analytics to help players enhance their winning chances through data-driven, actionable insights.
The central challenge in sports analytics lies in extracting meaningful insights from limited historical data to improve player performance. In racket sports such as tennis, discovering opponent-specific strategies is particularly critical. Although modern Large Language Models (LLMs) demonstrate strong general-purpose reasoning capabilities, our experiments show that their suggestions often lack the specificity and accuracy required for competitive sports strategy analysis. This limitation highlights the need for interpretable, data-driven approaches. To this end, we apply Probabilistic Model Checking (PMC), a formal method traditionally used in reliability analysis of complex systems such as aircraft, where the overall system reliability is assessed based on the reliability of its components (e.g., engines and sensors). In our context, a player's overall winning probability is modeled based on the reliability of sub-skills (e.g., serve, return, forehand). Using sensitivity analysis in PMC, our system can optimize strategies—for instance, recommending a 10% adjustment in a serving pattern to improve the winning probability against a specific opponent by 2%.
To achieve this, we utilize the Process Analysis Toolkit (PAT) model checker and the Probabilistic CSP# (PCSP#) modeling language for probabilistic reasoning. As part of this research, we formalized the syntax and denotational semantics of PCSP#, enabling PAT to model and verify concurrent, probabilistic systems effectively. This foundational work ensures the robustness of subsequent applications in sports analytics.
Building on this formal framework, we developed an interactive system for tennis strategy analytics using Markov Decision Processes (MDPs) and PMC. This system simulates tennis matches using historical data to evaluate strategies and provide interpretable, actionable recommendations for players and coaches. Evaluations using a decade of professional singles match data demonstrate the effectiveness of the system in improving decision-making and improving match outcomes. We further compare our strategy recommendations with those generated by LLMs, showing that our method consistently produces more accurate and opponent-specific insights. We extend this methodology to doubles matches by modeling team dynamics and cooperative strategies, such as player formations. Analysis of NCAA Division 1 doubles match data demonstrates the model's effectiveness in predicting match outcomes and providing tactical recommendations.
To make our analysis accessible to broader communities such as college and junior players, we developed a deep-learning framework for automated shot-by-shot data extraction from match videos. Addressing the fast, frequent, and fine-grained (F3) nature of sports events, we first introduce F3Set, a benchmark video dataset for F3 event detection. F3Set challenges existing temporal action understanding methods and advanced multimodal LLMs with over 1,000 event types at varying granularity levels. We also propose F3ED, a novel detection method that outperforms current approaches in detecting F3 events. This benchmark and method enable precise, multi-sport video analysis, offering insights for player evaluation and performance improvement. Furthermore, by introducing F3Set, we hope it can help advance multimodal LLM capabilities in F3 sports video understanding in the future. Moreover, we propose a multimodal distillation framework for few-shot temporal event detection, designed to tackle the challenge of data scarcity in fine-grained sports video understanding. The method leverages skeleton-based features to guide visual learning across RGB and optical flow streams through a three-stage training pipeline. Experiments on four sports datasets show that our approach achieves superior performance in few-shot settings, providing a scalable solution for fine-grained sports video analytics.
Extending the probabilistic reasoning framework beyond racket sports, we apply it to more complex team sports. We first develop an MDP-based strategy model for basketball, simulating half-court possessions using NBA player-tracking data and box score statistics. The model captures key decision-making processes such as passing, shooting, and rebounding. Evaluation against historical game outcomes and betting-based performance metrics demonstrates its predictive capability and practical utility for team-level strategy analysis. We further generalize this approach to American football by modeling NFL drive outcomes through PMC. Using historical play-by-play data, our model abstracts drive sequences into high-level decisions and quantifies the impact of strategic choices on scoring probabilities.
Finally, we consolidated these components into a web-based sports analytics platform that supports end-to-end workflows. This platform integrates video analysis, match outcome prediction, and opponent-specific strategy recommendations, offering tools such as player reports and pre-match decision support. Initially designed for tennis, the platform is adaptable to broader sports contexts, demonstrating the versatility and scalability of the proposed framework.