Using Machine Learning to Identify Key Kinematic Factors in Male Race Walkers at Different Competition Speeds

Against the evolving landscape of elite men’s race walking, where success is increasingly determined by the integration of speed and technique, this study establishes a machine learning-based framework to identify key kinematic factors across different competition speeds. Traditional analytical methods, often limited by laboratory conditions and linear models, struggle to capture the complex, nonlinear speed–technique relationships in real-world settings. To address this, kinematic and GPS data were collected from elite male race walkers under actual competition conditions. Using 17 candidate technical variables, multiple machine learning models were trained to classify three speed zones. The Extra Trees model demonstrated superior performance, indicating pronounced nonlinearity and high-dimensional interactions in the data. Interpretable AI techniques, including SHAP analysis and partial dependence plots, identified step frequency, flight distance/time, step length, and rear-pedal angle as the core determinants of speed classification. These factors exhibit distinct nonlinear effects and threshold ranges across speed zones, revealing a coupling structure where step frequency is central, while step length and flight variables act as constraints. Case-based explanations further highlighted individualised technical solutions within this shared structure. This research provides a novel, interpretable, and actionable “speed–technique” framework that can directly inform targeted training interventions, facilitating a shift from experience-based to data-driven training in elite race walking.