Public Sentiment Analysis of Indonesia’s Free Nutritious Meals (MBG) Program Using Lexicon-Based Labeling, SMOTE, and Machine Learning

Abstract

The Free Nutritious Meals (MBG) Program is a strategic policy of the Indonesian government aimed at improving national nutrition and has elicited diverse public reactions on social media. This study analyses public sentiment toward the MBG program using 11,003 cleaned public comments. Because the dataset lacked sentiment labels, a lexicon-based labelling approach using the Indonesian InSet Lexicon was applied to automatically classify comments as positive, negative, or neutral. The labelled data were then divided into 80% training data and 20% testing data using stratified sampling, and TF-IDF was used for feature extraction. Because the resulting class distribution was imbalanced (61.9% negative, 20.3% positive, and 17.8% neutral), the Synthetic Minority Over-sampling Technique (SMOTE) was applied only to the training data. Four classification algorithms (Naive Bayes, Support Vector Machine, Logistic Regression, and Random Forest) were trained and compared with and without SMOTE. The results show that SMOTE improved the macro-average F1-score across all models, particularly Naive Bayes (from 0.42 to 0.67). At the same time, SVM remained the best-performing model, achieving 84.83% accuracy and a macro F1-score of 0.81 after SMOTE, with a slight accuracy trade-off for substantially better recall in the neutral and positive minority classes. These findings indicate that combining lexicon-based labelling with SMOTE and classical machine learning provides a more balanced and reliable tool for monitoring public sentiment toward government policy programs such as MBG.
Keywords: sentiment analysis; Free Nutritious Meals; lexicon-based labeling; SMOTE; machine learning