Comparison of Supervised Machine Learning Methods for Predicting Stunting Prevalence in West Java Province

Abstract

Stunting is a condition affecting children's growth and development due to chronic malnutrition and recurring infections, characterized by a height below -2 standard deviations on the WHO growth curve. It remains a major global nutritional issue, with 149.2 million children (22%) affected worldwide in 2020. In the same year, 276,069 children in West Java (24.5%) were classified as stunted. Addressing this issue can involve predictive approaches, such as supervised machine learning. The methods compared include Polynomial Regression (PR), Support Vector Regression (SVR), and Linear Regression (LR) in four treatments. The models analyzed include PR, SVR, LR, PR-XGB (Polynomial Regression with XGBoost), SVR-SGB (Support Vector Regression with Stochastic Gradient Boosting), LR-XGB (Linear Regression with XGBoost), PR-XGB2, SVR-SGB2, LR-XGB2 (the previous models with double boosting), PR-XGB2-Opt, SVR-SGB2-Opt, and LR-XGB2-Opt (double boosting with hyperparameter optimization). The novelty of this study lies in improving the performance of the models through a double-boosting technique using Extreme Gradient Boosting (XGBoost) with hyperparameter optimization via GridSearchCV. Among models without boosting, LR achieved the best performance with MSE 0.018217, MAE 0.130036, MAPE 0.314071; with single boosting, SVR-XGB performed best with MSE 0.031485, MAE 0.162925, MAPE 0.344510; with double boosting and hyperparameter optimization, both models LR-XGB2 and LR-XGB2-Opt maintained the best performance with the same value, MSE 0.016474, MAE 0.124677, MAPE 0.309293. These results suggest that double boosting with proper tuning significantly enhances model performance in predicting stunting prevalence.