<![CDATA[Introduction: TikTok has become an important platform for public discussion of government programs, including Indonesia’s Free Nutritious Meal (MBG) Program, while the large volume of comments makes manual sentiment analysis inefficient. This study evaluates Multinomial Naïve Bayes for classifying public sentiment toward the MBG Program using TikTok comments. Method: A total of 1,048 comments were collected from the official National Nutrition Agency TikTok account using Apify, with 1,028 comments remaining after preprocessing. Sentiment pseudo-labels were generated using IndoRoBERTa, producing 333 positive and 695 negative comments. Undersampling balanced the dataset to 666 comments, followed by an 80:20 stratified train–test split. TF-IDF feature extraction and Multinomial Naïve Bayes classification were integrated in a Scikit-learn Pipeline, while GridSearchCV with 5-fold cross-validation optimized model parameters. Results and Discussion: The optimized model achieved 80.60% accuracy, 81.04% precision, 80.60% recall, and an 80.53% F1-score on 134 test comments. Negative sentiment dominated the original post-preprocessing dataset at 67.61%. However, these results primarily indicate agreement with IndoRoBERTa-generated pseudo-labels rather than human-verified sentiment labels. Conclusion: The TF-IDF–Multinomial Naïve Bayes pipeline provides a useful baseline for MBG-related TikTok sentiment classification, but manual label validation and comparison with alternative classifiers are required to establish stronger reliability and generalizability.]]>
