Multilabel Aspect-Based Sentiment Analysis pada Ulasan Halodoc Menggunakan Binary Relevance dan SVM

Abstract

User reviews on the Halodoc telemedicine application are an important data source for evaluating service quality, but often discuss more than one service aspect with different sentiments in a single review. This condition makes single-label sentiment classification less optimal. This study implements Multilabel Aspect-Based Sentiment Analysis on 2,390 Indonesian-language Halodoc reviews collected from Google Play Store during January to December 2025. Each review was manually labeled by three independent annotators using eight binary labels across four service aspects, combined through majority voting. Data was split 80:20 into 1,912 training samples and 478 testing samples, features were extracted using TF-IDF, and modeling used Onevsrestclassifier(LinearSVC) with a Binary Relevance strategy. Hyperparameter testing via 5-fold cross-validation identified C=100 as the best parameter with a macro Avg F1-score of 0.5672. Final model evaluation on the test data produced a Hamming loss of 0.0662, Subset accuracy of 0.6402, micro avg F1-score of 0.8046, and macro avg F1-score of 0.7179. The results show the highest positive sentiment proportion on Application Performance at 0.7412 and Service Quality at 0.7138, while the highest negative sentiment proportion appears on Price at 0.9189 and Service Completeness at 0.6279.