A Study on Predicting Adolescent Cross-Sport Performance Using Shared Feature Space and Multi-Task Regression

Abstract

Common transfer patterns exist across different sports disciplines in adolescents regarding abilities such as speed, explosiveness, and rhythm control. This study constructs a cross-discipline prediction framework based on shared feature space, targeting throwing velocity, sprint performance, and jumping ability. The framework comprises a general ability feature layer, discipline-specific feature layer, and multi-task regression head. General features include motion rhythm statistics, velocity distribution, acceleration trends, and training load indicators, while the discipline layer incorporates discipline-specific motion information. Predictions are performed using multi-head LightGBM. Experiments on 2.1 million sequence data points from 619 athletes demonstrate an average MAPE of 6.5%, representing a 22.7% improvement over single-task models. The framework maintains stable advantages even under sparse data conditions, indicating that the shared feature structure effectively enhances cross-sport prediction capabilities for adolescent athletic performance.