Sentiment Analysis Mobile JKN Reviews Using SMOTE Based LSTM

Abstract

The JKN Mobile application plays an important role in providing easy and fast access to health services for JKN-KIS users. However, user reviews indicate dissatisfaction with several aspects of the application, such as login issues and OTP codes, which can affect the overall user experience. Another challenge faced is class imbalance in the review dataset, which can affect the performance of sentiment analysis. This study uses Long Short-Term Memory (LSTM) combined with Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance. Review data was collected from Google Play Store and Kaggle, then preprocessed including lemmatization, tokenization, and padding. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics. The results showed that LSTM with SMOTE achieved 88% accuracy, 90% precision, 88% recall, and 89% F1-score. SMOTE successfully improved performance in the minority class although there was a slight decrease in accuracy compared to the model without SMOTE. Word cloud visualization reveals positive sentiments regarding the ease of use of the application, while negative sentiments indicate areas that need improvement. This study emphasizes the importance of handling imbalanced datasets to produce more accurate sentiment analysis.