ChampionNet: a transformer-enhanced neural architecture search framework for athletic performance prediction and training optimization

Abstract Neural architecture search (NAS) has emerged as a promising approach for automating deep learning model design. However, its application in sports analytics faces unique challenges due to the complex interplay between biomechanical patterns, physiological adaptations, and coaching expertise. Traditional NAS methods need help to effectively capture the multifaceted nature of athletic performance, often failing to integrate qualitative coaching insights with quantitative measurements. We introduce ChampionNet, a framework incorporating NAS and large language models to enhance accuracy in predicting athletic performance and tailoring training regimens. Our approach offers three primary contributions: integrating hyperdimensional embedding to capture fine-grained biomechanical features and physiological parameters with exceptional detail, a structure-preserving graph encoding leverages to maintain crucial spatiotemporal relationships in athletic movements, and the novel comprehensiveness of the training graph that models forward performance prediction and backward physiological adaptation pathways. Our experiments on various sports demonstrate that ChampionNet outperforms other models by 2.5% in accuracy and over 61.9% in computational cost. Further insights illustrate the framework's performance with complex patterns and multi-modal data, especially for sports with advanced biomechanical needs. These findings support ChampionNet's effectiveness as an integrative athletic performance optimization solution, highlighting the need for automated architecture search tailored to sports.