Intelligent Assessment of Athlete Physical Fitness: Addressing Data Imbalance

Abstract

This study aims to mitigate the impact of imbalanced data through the use of the oversampling technique and to develop supervised learning models for assessing the physical fitness of youth athletes. The dataset comprises the physical fitness test results from 75 athletes aged 11 to 16 years. The dataset presents two major challenges: A limited sample size and a significant class imbalance, with certain fitness levels being underrepresented. This class imbalance can substantially degrade the performance of classification models, as it often leads to biased predictions favoring the majority class while failing to learn the characteristics of minority classes, those that may be most critical in practice. To address this issue, the Synthetic Minority Oversampling Technique (SMOTE) was employed to synthetically balance the class distribution. Five supervised learning algorithms were evaluated: Light Gradient Boosting Machine, Decision Tree, Random Forest, Neural Network, and Multinomial Logistic Regression. The Light Gradient Boosting Machine model yielded the highest accuracy at 87.76%, followed by Decision Tree, Random Forest, and Neural Network models, each with an accuracy of 79.59%. The Multinomial Logistic Regression model achieved the lowest accuracy at 75.51%. On average, the classification accuracy across all models improved to 81.41%, representing a 12.23% increase compared to using the original imbalanced dataset. The results demonstrate that applying oversampling techniques such as SMOTE can effectively alleviate the effects of class imbalance and enhance the predictive performance of machine learning models in the context of physical fitness assessment.