Explainable Sentiment Analysis pada Tweet di X: Perbandingan antara Pendekatan Transformer dan Machine Learning Klasik

Abstract

Sentiment analysis in political discourse on social media is crucial for understanding trends in public opinion. However, in the Indonesian context, comparative studies between classical models that include explainability analysis and transformer-based models are still limited. This study compares the performance and decision making patterns in SVM & TF-IDF and IndoBERTweet in binary sentiment classification, using tweets about the Indonesian Minister of Finance, Purbaya Yudhi Sadewa. Using 3.520 tweets in Indonesian collected from August to October 2025, a dual preprocessing strategy was applied, aggressive normalization with negative labeling for SVM and minimal intervention for IndoBERTweet. To address the class imbalance ratio of 4.28:1, the models were trained using temporal split with class weighting. IndoBERTweet significantly outperformed SVM in all metrics, showing a very significant improvement in F1-Macro and minority class detection. This was confirmed by a significant effect size through the McNemar test (p= 0.0021). This advantage focuses on detecting negative sentiment, although minority class classification remains a challenge in conditions of extreme temporal imbalance. Explainability analysis based on SHAP shows significant differences between the models without any overlap between their features. While SVM relies on domain specific keywords that reflect concrete entities and policies, IndoBERTweet uses evaluative verbs and modal construction that reflect situational context and discourse patterns. Issue salience estimation identified Governance and Leadership, Law and Regulation, and Economic and Fiscal Policy as the most salient issues. This is in line with the discourse that developed in October 2025 discourse dominated by Coretax implementation and the rejection of using the state budget to pay off the Whoosh high speed railway debt.