This study proposes a big data-driven optimisation system integrating graph neural networks (GNNs) and ensemble learning to overcome the limitations of traditional athlete selection and training methods with high-dimensional, multi-modal data.The system constructs a heterogeneous graph with athletes as nodes and relationships as edges.GNNs exploit complex associations among multi-dimensional features.A novel stacked ensemble framework, incorporating XGBoost and random forest, enhances model generalisation and robustness.This creates a closed loop from static evaluation to dynamic training optimisation, enabling personalised training plans based on real-time data.Experiments on a multi-source dataset of 1,000 athletes (fitness tests, performance, and physiological indicators) show the system's potential prediction accuracy reaches 94.5%.The core GNN + Stacking module achieves 90.5% accuracy on basic feature subsets -about 12% higher than traditional models -and can reduce sports injury risk by approximately 18%.
Mu et al. (Thu,) studied this question.