Knowledge-grounded large language model for personalized sports training plan generation

Abstract

The growing demand for scientifically grounded and highly personalized fitness plans reveals the huge shortcomings of traditional recommender systems, which cannot overcome template-oriented methods and effectively cope with complex, dynamic user data. As a remedy for this shortcoming, this work utilizes a Large Language Model (LLM) augmented with a domain-specific knowledge graph to develop LLM-SPTRec, a novel framework for intelligent sports training plan generation. This model successfully integrates multi-source heterogeneous user data and enhances the personalization and scientific validity of recommendations by grounding the LLM’s generative process in an expert-elicited Sports Science Knowledge Graph (SSKG). Empirical results on a real-world dataset demonstrate that LLM-SPTRec surpasses traditional baselines—including collaborative filtering, sequential models, and general-purpose LLMs—on fundamental measures of plan coherence, goal relevance, and predicted user satisfaction. The findings of this research provide a new paradigm for the discipline of intelligent health by bridging the gap between big data analysis and expert knowledge in addition to providing a new direction for the overall field of applied AI by demonstrating that knowledge-based LLMs are capable of generating safe, effective, and scientific personal health recommendations.