AI-based personalized sports training plan generation system: A multi-community ranking framework

We propose a scalable hybrid multi-modal framework for personalized sports training plan generation based on diverse athlete profiling and dynamic optimization search. Designed for large-scale athletic environments, our system uncovers coherent, performance-aligned athlete communities through hierarchical feature learning. This pipeline extracts diverse biometric features—such as cardiovascular capacity, movement efficiency, and recovery patterns—enhanced by leveraging dynamic programming to identify optimal training paths for each athlete. A physiological feature selection strategy filters out less discriminative attributes, producing refined performance representations. These athlete representations are projected into a latent performance space, where each individual is modeled as a probabilistic distribution over abstract training themes, supporting precise ability differentiation. We construct a weighted athlete similarity graph from these representations, enabling large-scale community detection through advanced clustering techniques. The resulting training communities reflect shared physiological patterns—such as endurance-focused or power-oriented athletes—and reveal both macro- and micro-level performance trends. To deliver personalized training plans, a ranking module integrates individual athlete profiles with community embeddings to suggest optimal exercise regimens. This fusion of individual characteristics and group patterns enhances training effectiveness. Evaluations on a dataset of over one million training sessions confirm the system's scalability and prediction accuracy, demonstrating robustness across varied fitness levels and effectiveness in large-scale personalization scenarios.