<![CDATA[The rapid adoption of digital banking generates a large volume of unstructured user reviews, making document-level sentiment analysis insufficient to identify the specific sources of satisfaction and dissatisfaction.This study applies Aspect-Based Sentiment Analysis (ABSA) to user reviews of BRImo and Livin’ by Mandiri by comparing TF-IDF and Word2Vec representations using Support Vector Machine (SVM). Reviews were collected from Google Play Store with google-play-scraper using newest-first ordering and a target of up to 30,000 reviews per application.After cleaning and selection, 41,461 clean reviews remained, and sentiment classes were balanced to 8,943 records for modeling. Aspect extraction used a dictionary-based approach, while three sentiment classes were assigned from star ratings.TF-IDF used 1–2 grams with a maximum of 5,000 features, whereas Word2Vec used a 100-dimensional Skip-gram model with average pooling.SVM hyperparameters were optimized using Grid Search and 5-Fold Cross-Validation. Both feature schemes achieved 64.6% accuracy, 66.7% macro precision, 64.6% macro recall, and 65.3% macro F1-score. BRImo was dominated by positive sentiment (64.1%), whereas Livin’ by Mandiri was dominated by negative sentiment (49.2%), with Login as the most prominent issue. The findings indicate that TF-IDF and Word2Vec were practically equivalent on this dataset,while ABSA provides actionable evidence for prioritizing digital banking service improvements]]>
Copyrights © 2026
