Comparative Analysis of SMOTE, WMOTE, and ADASYN Oversampling Methods on Multinomial Naive Bayes Performance in Classifying Toddlers Nutritional Status

Class imbalance in toddler nutritional status data often reduces the ability of classification models, especially in predicting minority classes. This study aims to analyze the impact of three oversampling techniques, namely SMOTE, WMOTE, and ADASYN, on improving the performance of the Multinomial Naive Bayes (MNB) algorithm. A dataset of 243 data was processed through a preprocessing stage by converting categorical variables using numeric labels. To meet the MNB algorithm's requirement for non-negative data, continuous numeric features (such as birth weight, birth height, weight, height, and age) were normalized using the Min-Max Scaler to the range [0, 1]. This process discretizes continuous values onto a probability scale to ensure feature compatibility with the Multinomial distribution. Data balancing was performed only on the training dataset, where the SMOTE method produced 374 data, ADASYN produced 375 data, and WMOTE produced 373 data. The evaluation results show that although all three oversampling methods experienced a slight decrease in global accuracy, the model's ability to detect minority classes improved, as evidenced by increases in G-Means and Balanced Accuracy. The test results concluded that MNB-ADASYN was the best model for prioritizing high sensitivity to all class labels, while MNB-WMOTE provided the most consistent global accuracy stability while maintaining performance on minority classes.