Aspect-Based Sentiment Analysis on User Perceptions of OVO using Latent Dirichlet Allocation and Support Vector Machine

The rapid development of digital technology and the Internet has significantly influenced financial services in Indonesia, leading to the widespread use of digital wallets. One of the most prominent digital wallet platforms is OVO, which has received millions of user reviews across application stores. This study applies aspect-based sentiment analysis to better understand user perceptions from reviews of the OVO application (versions 3.115 to 3.119). A total of 17.086 reviews were collected through web scraping and refined to 4.996 relevant entries. Topic modeling using Latent Dirichlet Allocation (LDA) identified four main aspects frequently discussed by users: Transaction Efficiency, User Experience, Account Access and Registration, and Balance and Charges. However, automatic aspect labeling using LDA keywords achieved only 11.46% agreement with manual annotations, increasing to 40.60% after keyword refinement. Therefore, manual aspect annotation was adopted as the basis for sentiment labeling. Sentiment labeling was conducted by three annotators based on structured guidelines, achieving a Fleiss’ Kappa score of 0.9915. A classification model was then developed using the Support Vector Machine (SVM) algorithm across six testing scenarios. The best-performing model, using a Linear kernel without ML-SMOTE, achieved a macro-average precision of 0.843, recall of 0.786, and F1-Score of 0.804. These results demonstrate the model’s effectiveness in handling multi-label classification under imbalanced data conditions, particularly for well-distributed aspects such as Transaction Efficiency and User Experience, while highlighting challenges in minority-class detection for aspects such as Account Access and Registration and Balance and Charges.