Comprehensive Analysis of TF-IDF and Word2Vec-Based Text Feature Representation on Support Vector Machine Performance in Sentiment Analysis of Telemedicine Reviews

Abstract

User reviews of telemedicine applications provide valuable insights into user satisfaction and service quality. Accurate sentiment analysis of these reviews depends greatly on the text representation method used before classification. Although TF-IDF and Word2Vec have been widely applied in sentiment analysis, their comparative performance on Indonesian-language telemedicine reviews with a highly imbalanced sentiment distribution has not been sufficiently investigated. So, in order to analyze the sentiment of Halodoc application evaluations using the Support Vector Machine (SVM) classifier, this study analyzes the performance of TF-IDF and Word2Vec, two text representation approaches. The Google Play Store was combined from January 2024 to October 2025 for a grand total of 10,443 reviews. After implementing user-rated automatic sentiment labeling, 6,574 valid reviews were obtained using Indonesian-specific preprocessing that included case folding, cleaning, stopword removal, and stemming. Stratified 5-fold cross-validation using recall, accuracy, precision, and F1-score metrics was used to test the models. Word2Vec only managed 75% accuracy, 78% precision, 79% recall, and 78% F1-score in the experiments, but TF-IDF achieved 90% accuracy, 89% precision, 90% recall, and 89% F1-score. TF-IDF outperformed Word2Vec because frequency-based features are more effective in representing short review texts, whereas Word2Vec generally requires a substantially larger corpus to learn robust semantic representations. These findings provide empirical evidence for selecting appropriate text representation methods in Indonesian telemedicine sentiment analysis and contribute practical guidance for developing more accurate digital healthcare opinion mining systems.