<![CDATA[Important information regarding user satisfaction and complaints about an application can be obtained from reviews uploaded to the Google Play Store. Telecommunications service applications are among the most widely used applications in Indonesia and receive numerous reviews that can be leveraged to evaluate service quality. However, manually analyzing large volumes of reviews is inefficient. This study aims to classify sentiment in telecommunication application user reviews by comparing TF-IDF Unigram (Baseline) and N-Gram (Unigram-Bigram) feature representations and evaluating Naive Bayes and Support Vector Machine (SVM), with hyperparameter validation using Grid Search applied separately to each feature scenario. The dataset consists of 2,836 user reviews collected from the Google Play Store through web scraping. The data underwent preprocessing, rating-based sentiment labeling with rule-based label correction, data splitting, and TF-IDF feature extraction for both scenarios before classification. Model performance was evaluated using accuracy, precision, recall, F1-score, F1-score macro average, and confusion matrix. The results show that SVM outperforms Naive Bayes in both feature scenarios. The best performance was achieved by SVM using the Baseline TF-IDF Unigram representation, with 92.43% accuracy, 91.98% precision, 88.24% recall, 90.07% F1-score, and 91.98% F1-score macro average. Grid Search produced identical optimal configurations for both scenarios, namely linear kernel, C=1, and gamma=scale, indicating their suitability for both feature representations. These findings indicate that, within the dataset and experimental configuration used, adding N-Gram features does not necessarily improve classification performance, particularly for SVM, while unigram representation alone effectively classifies telecommunication application user reviews.]]>
Copyrights © 2026
