Sentiment Analysis of LinkAja Digital Wallet Application Reviews on Google Play Store using Transfer Learning IndoBERT

Abstract

The LinkAja digital wallet receives an average rating of 3.5 on the Google Play Store despite having a higher number of user reviews than its competitors, indicating a strong need for data-driven evaluation of user satisfaction. This study performs sentiment classification on LinkAja user reviews using the IndoBERT model implemented within the CRISP-DM framework. A total of 1,483 reviews posted from January 1 to May 31, 2025, were analyzed through automatic labeling using a pretrained IndoBERT sentiment model and validated using an 80:20 hold-out scheme. Model performance was evaluated using accuracy, the F1 score, and the Matthews Correlation Coefficient (MCC) to address class imbalance. The results show high classification performance with 95% accuracy, a macro F1-score of 0.92, a weighted F1-score of 0.94, and an MCC of 0.90. Sentiment distribution reveals a dominance of negative sentiments at 59.5%, followed by positive (26.1%) and neutral (14.4%) sentiments. Theoretically, this study reinforces the superiority of IndoBERT over conventional machine learning methods for Indonesian sentiment analysis. Practically, the findings provide actionable insights into service improvements, particularly regarding transaction stability and system reliability.