<![CDATA[Abstrak Penelitian ini bertujuan untuk membandingkan performa tiga pendekatan analisis sentimen pada layanan IndiHome menggunakan data Twitter, yaitu Naïve Bayes, Support Vector Machine (SVM), dan Indonesian Bidirectional Encoder Representations from Transformers (IndoBERT). Keterbatasan model tradisional melatarbelakangi penelitian ini dalam mengenali opini positif dan tantangan ketidakseimbangan data yang sering muncul dalam analisis berbasis media sosial. Data penelitian berupa 7393 tweet (Januari 2019–Agustus 2024) yang dilabeli secara manual menjadi sentimen positif dan negatif. Model dievaluasi menggunakan stratified 5-fold cross validation dan data uji, dengan penerapan teknik penanganan ketidakseimbangan berupa Synthetic Minority Oversampling Technique (SMOTE) dan pembobotan kelas (class weighting). Hasil menunjukkan IndoBERT unggul dengan akurasi 0,96 dan F1-score makro 0,95 tanpa penanganan khusus, sedangkan SVM mencapai akurasi 0,95 dengan pembobotan kelas, dan Naïve Bayes meningkat dari akurasi 0,89 menjadi 0,92 setelah SMOTE. Analisis tren sentimen menunjukkan opini negatif mendominasi, terutama terkait kecepatan dan kestabilan layanan. Temuan ini menegaskan bahwa IndoBERT lebih efektif dalam memahami konteks bahasa Indonesia, sementara teknik penanganan data tetap relevan untuk meningkatkan performa model tradisional. Kata Kunci: Analisis Sentimen, Stratified 5-Fold Cross Validation, SMOTE, Pembobotan Kelas. Abstract This study aims to compare the performance of three sentiment analysis approaches on IndiHome services using Twitter data, namely Naïve Bayes, Support Vector Machine (SVM), and Indonesian Bidirectional Encoder Representations from Transformers (IndoBERT). The limitations of traditional models drive the background of this research in recognizing positive opinions and the challenge of data imbalance that often arises in social media-based analysis. The research data consists of 7393 tweets (January 2019–August 2024) manually labeled into positive and negative sentiments. The model is evaluated using stratified 5-fold cross-validation and test data, with the application of imbalance handling techniques such as Synthetic Minority Oversampling Technique (SMOTE) and class weighting. The results show IndoBERT excels with an accuracy of 0.96 and a macro F1-score of 0.95 without special handling. At the same time, SVM achieves an accuracy of 0.95 with class weighting, and Naïve Bayes improves from 0.89 to 0.92 after SMOTE. Sentiment trend analysis shows a predominance of negative opinions, particularly regarding service speed and stability. These findings confirm that IndoBERT is more effective at understanding the Indonesian context, while data handling techniques remain relevant for improving the performance of traditional models. Keywords: Sentiment Analysis, Stratified 5-Fold Cross Validation, SMOTE, Class Weighting.]]>
