<![CDATA[Class imbalance is a recurring obstacle in machine learning based screening of child nutritional status. This study evaluates and compares the effect of SMOTE and SMOTE+Tomek Links on the classification of toddler nutritional status using K-Nearest Neighbours (KNN) and Random Forest (RF). The data consist of 9,426 anthropometric records with four predictors, labelled into three classes: normal (7,212; 76.51%), moderate stunting (1,612; 17.10%) and severe stunting (602; 6.39%), a majority to minority ratio of about 12:1. Min-Max scaling and resampling were fitted on training data only, inside a pipeline, and the models were assessed on an independent 20% test set (n = 1,886) with stratified 5-fold cross validation. At baseline, RF reached 0.975 accuracy and 0.932 macro-F1, while KNN displayed an accuracy paradox: 0.913 accuracy but only 0.392 recall on severe stunting (47 of 120 cases). SMOTE raised KNN recall to 0.733 and macro-F1 from 0.753 to 0.816. For RF, SMOTE improved both criteria at once: recall rose from 0.825 to 0.908 (99 to 109 cases) and macro-F1 from 0.932 to 0.949. SMOTE+Tomek performed almost identically, removing only 30 of 17,307 training samples. RF with SMOTE is therefore recommended]]>
Copyrights © 2026
