<![CDATA[Public opinion plays a crucial role in the successful implementation of renewable energy policies, particularly regarding the transition to ethanol-based fuels in Indonesia. Understanding this sentiment is vital to mitigate social resistance and design effective communication strategies, as policy failure often stems from public rejection rather than technical issues. However, social media data regarding this topic is often highly imbalanced, with a dominance of non-positive sentiments (96%) compared to positive ones (4%), creating a severe bias in machine learning models known as the accuracy paradox. This study aims to classify public sentiment towards ethanol policy and evaluate the effectiveness of the Synthetic Minority Over-sampling Technique (SMOTE) in handling extreme class imbalance. The methods used include text preprocessing with Sastrawi, feature extraction using TF-IDF, and a comparative classification between Logistic Regression (LR) and Random Forest (RF). The novelty of this research lies in addressing the extreme imbalance in high-dimensional text data, proving that simpler linear models can outperform complex ensemble models in terms of minority class detection. The results show that the Optimized Logistic Regression model with SMOTE outperformed Random Forest, achieving a Precision of 1 and an F1-Score of 0.67 for the minority class, compared to RF which only reached an F1-Score of 0.55. This study concludes that for high-dimensional sparse text data, linear models combined with SMOTE provide superior performance in identifying minority sentiments.]]>
