Analisis Sentimen Ulasan IKD Berbasis Lexicon dan SMOTE pada Naive Bayes, SVM, dan Random Forest

Abstrak

The rapid adoption of mobile-based public services in Indonesia has driven increased utilization of the Identitas Kependudukan Digital (IKD) application for population administration. As the user base expands, review data from the Google Play Store has grown substantially, reflecting varied user experiences encompassing service satisfaction, usability perceptions, and technical obstacles. Processing this volume of unstructured text manually is impractical, necessitating an automated computational approach to opinion mining. This study investigates opinion polarity in IKD user reviews through a lexicon-based labeling strategy and evaluates the classification capacity of Naive Bayes, Support Vector Machine (SVM), and Random Forest algorithms both prior to and following the application of Synthetic Minority Oversampling Technique (SMOTE). A total of 4,896 reviews were gathered from Google Play Store, with 3,708 categorized as negative and 1,188 as positive sentiments. The pipeline encompassed text preprocessing, automatic lexicon-driven labeling, TF-IDF vectorization, an 80:20 train-test partition, and multi-metric model evaluation. Experimental findings confirm that SMOTE effectively addresses class imbalance by bolstering the representation of the minority class. Among all configurations tested, SVM combined with SMOTE yielded superior results, recording 98.98% accuracy, 98.29% precision, 97.46% recall, and 97.87% F1-score. These outcomes demonstrate that integrating lexicon-based annotation, TF-IDF feature weighting, and SMOTE resampling constitutes an effective pipeline for sentiment classification in digital government applications, offering a scalable mechanism for evidence-based public service quality evaluation.