Comparing Machine Learning and IndoBERT for PPKM Twitter Sentiment Classification

During the COVID-19 pandemic, Indonesia's PPKM policy generated broad discussion on Twitter/X, making sentiment analysis useful for understanding public responses. Indonesian tweets are informal, contain spelling variation and negation, and have imbalanced labels. This study compares Random Forest, XGBoost, and fine-tuned IndoBERT for three-class sentiment classification of PPKM-related tweets. For the classical models, preprocessing preserves important negation markers and represents text with TF-IDF. Because Neutral tweets dominate, Macro F1-score is used as the main metric alongside accuracy. IndoBERT achieves the best performance, with 91.27% accuracy and a Macro F1-score of 0.8429. Random Forest obtains Macro F1-score 0.6664, slightly above XGBoost at 0.6620. Execution time is reported descriptively because the recorded stages are not identical. Kebijakan PPKM selama pandemi COVID-19 memicu diskusi luas di Twitter/X, sehingga analisis sentimen diperlukan untuk memahami respons masyarakat. Cuitan berbahasa Indonesia bersifat informal, memiliki variasi ejaan dan negasi, serta distribusi label tidak seimbang. Penelitian ini membandingkan Random Forest, XGBoost, dan IndoBERT yang di-fine-tuning untuk klasifikasi sentimen tiga kelas pada cuitan tentang PPKM. Pada model klasik, prapemrosesan mempertahankan kata negasi penting dan teks direpresentasikan dengan TF-IDF. Karena kelas Netral mendominasi, Macro F1-score digunakan sebagai ukuran utama selain akurasi. IndoBERT memberikan kinerja terbaik dengan akurasi 91,27% dan Macro F1-score 0,8429. Random Forest memperoleh Macro F1-score 0,6664, sedikit di atas XGBoost 0,6620. Waktu eksekusi dilaporkan secara deskriptif karena tahapan yang tercatat tidak identik.