Predicting Body Fat with Simple Physical Attributes

This study tackles the challenge of developing effective machine learning models for small datasets, a critical need in scenarios where collecting large datasets is impractical or infeasible. Using the Body Fat Prediction Dataset with 252 samples, we explore how simple anatomical measurements can deliver reliable predictions with minimal accuracy loss. We evaluate the impact of outlier filtering, feature engineering, predictor selection, and validation techniques across multiple regression models. Our findings show that lightweight models, such as linear regression and Bayesian ridge regression, achieve optimal performance with RMSE = 4.283%, surpassing traditional methods. In contrast, advanced models like Gradient Boosting (GB) show benefits only with larger datasets, highlighting the limited advantage of complexity in data-scarce scenarios. Abdominal circumference emerged as a key predictor, while other features, such as knee circumference, negatively impacted predictions.