Comparative Sentiment Analysis of YouTube Comments on Indonesia's Electric Vehicle Incentive Policy Using TF-IDF and IndoBERTweet Models

Abstract

Indonesia’s battery electric vehicle (KBLBB) incentives aim to accelerate low-carbon mobility, yet public reactions regarding affordability, charging infrastructure readiness, and subsidy equity remain highly heterogeneous. This research systematically compares classical machine learning and transformer-based models for classifying sentiment in 1,516 YouTube comments discussing the incentive policy and broader EV ecosystem. Comments are collected via web scraping and processed through filtering, case folding, normalization, tokenization, stopword removal, stemming, lexicon-based sentiment labelling, TF-IDF bigram vectorization, random oversampling, and hyperparameter optimization with GridSearch. Support Vector Machine and Random Forest serve as baseline models, while Logistic Regression with TF-IDF bigram and IndoBERTweet represent advanced approaches that exploit richer feature representations. Results show that the baseline models achieve around 65–66% accuracy, Logistic Regression improves performance to 88%, and IndoBERTweet attains the highest accuracy of 94% with balanced precision, recall, and F1-score across sentiment classes. Sentiment distribution indicates that 63.3% of comments are negative, dominated by concerns over limited charging networks, high upfront costs, and perceived unfairness of public subsidies, while 36.7% of comments highlight support for cleaner transportation, technological innovation, and national industrial competitiveness. These findings demonstrate that transformer-based contextual embeddings substantially enhance sentiment classification for noisy Indonesian social media text and provide a scalable informatics tool for continuous monitoring of EV policy reception. The resulting empirical evidence can inform more targeted refinements of incentive design, infrastructure planning, and communication strategies, thereby supporting inclusive, data-driven, and sustainable KBLBB adoption across diverse demographic groups and evolving policy scenarios nationwide over time.