Hybrid Cost-Sensitive Rejection Sampling Approach for Improving Stunting Classification Performance on Imbalanced Data

Abstract

Stunting is a serious child health problem in Indonesia, with a national prevalence of 21.6% in 2022. Developing machine learning-based classification models for stunting detection faces a major challenge due to class imbalance, where stunting cases are far fewer than non-stunting cases. This study evaluates six decision tree-based classification algorithms (J48Consolidate, CDT, C4.5, LADTree, LMT, and Hoeffding Tree) combined with five imbalance handling techniques: SMOTE, undersampling, oversampling, Cost-Sensitive Oversampling, and Cost-Sensitive Rejection Sampling (CSRS), using a dataset of 10,000 records with an imbalance ratio of 3.89:1. Results show that CSRS consistently achieved the best performance across all sampling methods. The optimal combination of C4.5 and CSRS yielded a Precision of 0.963, Recall of 0.970, F-Measure of 0.960, and Accuracy of 0.9703. CSRS outperformed other methods by working exclusively with real samples, avoiding the risk of clinically unrepresentative synthetic data.