MULTI-CLASS SENTIMENT ANALYSIS USER REVIEWS OF THE HALODOC APPLICATION USING LONG SHORT-TERM MEMORY (LSTM)
Eddy Prasetyo Nugroho, Alya Chairunnisa Faz, Ani Anisyah, Erna Piantari, Enjun Junaeti

Universitas Pendidikan Indonesia

Abstract

In the digital era, healthcare service applications like Halodoc have gained significant popularity among the public, and there is undoubtedly a wealth of user feedback available. This feedback is crucial for assessing the service and gaining insights into user satisfaction. Through sentiment analysis, users^ perspectives on the effectiveness of the service and its features can be better understood. LSTM method is effective in addressing long-term dependencies in sequential data, making it suitable for text analysis that requires an understanding of context and the interrelatedness of information in user reviews. Therefore, this research aims to analyze the sentiment of user reviews of the Halodoc application on the Google Play Store and App Store using the Long Short-Term Memory (LSTM) model. Sentiment analysis in this study involves two labels: the first label contains positive and negative classes, while the second label includes eight classes, including fast, helpful, reliable, professional, unprofessional, complicated, complex, and slow. The model^s performance shows variations depending on the number of epochs. In binary cross-entropy, accuracy reaches 99.89%, F1-Score achieves 90.11%, recall is at 93.18%, and precision stands at 87.23% after 50 epochs. Meanwhile, in categorical cross-entropy, the results are slightly lower, with accuracy reaching 98%, F1-Score at 90%, recall at 92%, and precision at 87% after 50 epochs. The LSTM model effectively addresses sentiment analysis of reviews with an understanding of context and long-term data relationships.

Keywords: Halodoc, Long Short-Term Memory, Multi-Class, Sentiment Analysis

Topic: Computer Science