Comparison of classical machine learning methods in the task of obesity level classification

Obesity is a critical public-health problem amenable to data-driven risk assessment. We compare eight classical classifiers on the UCI ObesityDataSet (2087 samples, 7 classes) under a leakage-free protocol: stratified 10-fold cross-validation, randomised-search tuning, pairwise McNemar and Wilcoxon tests with Holm–Bonferroni correction, and multi-method interpretability (split gain, SHAP, permutation, LIME). XGBoost reached 96.17 % test accuracy (F1 = 0.962, κ = 0.955); paired tests show no significant edge over tuned LightGBM, SVM-RBF, Random Forest or Logistic Regression. All four importance measures rank body weight and height first, and LIME recovers BMI-based thresholds, supporting gradient boosting for clinical decision support.