<![CDATA[Sentiment analysis of public policies on social media is crucial for government evaluation; however, it is often challenged by highly imbalanced datasets. This study aims to address this issue through a case study on public sentiment toward bioethanol fuel policies on YouTube, where the cleaned dataset after preprocessing consisted of 2,409 comments dominated by negative sentiment (1,430 comments), followed by neutral sentiment (734 comments), and only a small number of positive sentiments (245 comments). The performance of classical Machine Learning (ML) models was severely degraded due to this imbalance, particularly in detecting the minority class. This study applied TF-IDF weighting for feature extraction, followed by the Synthetic Minority Oversampling Technique (SMOTE) to balance the training data (1,927 samples) before comparing the performance of three ML algorithms: Logistic Regression, Support Vector Machine (SVM), and LightGBM. The evaluation results on the testing dataset (482 samples) demonstrate that the implementation of SMOTE significantly improved the models’ ability to recognize the “Positive” class. The LightGBM model combined with SMOTE achieved the best performance, with an accuracy of 64.11%. In particular, the application of SMOTE successfully increased the minority-class F1-score from a baseline of 18.18% to 35.29%. These findings confirm that handling imbalanced data is a critical step in producing valid and reliable sentiment analysis results.]]>
