<![CDATA[This study addresses a critical gap in automated public service evaluation by systematically comparing the performance of XGBoost and Support Vector Machine (SVM) for sentiment classification of BPJS Kesehatan application reviews. Unlike prior research that predominantly relies on default model configurations or single-algorithm frameworks, this study introduces a rigorously optimized comparative pipeline using GridSearchCV with k-fold cross-validation, specifically designed to address hyperparameter sensitivity and class imbalance in Indonesian digital health feedback. User reviews were extracted from the Google Play Store, preprocessed using a standardized NLP pipeline, and vectorized via TF-IDF. Analytical results reveal that while SVM achieves marginally higher overall accuracy (90.5%) through optimal hyperplane separation, it completely fails to classify neutral sentiments (F1-score = 0.00), highlighting its vulnerability to minority-class underrepresentation. In contrast, XGBoost (89.75% accuracy) demonstrates superior multi-class equilibrium, leveraging ensemble regularization to effectively capture ambiguous and neutral expressions. The systematic integration of GridSearchCV significantly improves generalization, validating hyperparameter optimization as a critical determinant of model robustness in real-world textual data. Scientifically, this study advances methodological understanding by demonstrating the trade-offs between margin-based strictness and ensemble adaptability under exhaustive optimization, providing a reproducible framework for imbalanced sentiment classification. Practically, it offers public health administrators a scalable, data-driven mechanism for real-time service quality monitoring and user satisfaction analytics.Keywords: Sentiment Analysis; XGBoost; Support Vector Machine]]>
Copyrights © 2026
