Pemanfaatan Bi-LSTM dan GloVe dengan SMOTE untuk Menganalisis Sentimen Pengguna Bank Digital

Abstract

The rapid growth of digital banking in Indonesia has generated a large volume of user reviews on platforms such as Google Play Store, which contain valuable feedback regarding service quality. This study aims to perform sentiment analysis on digital banking user reviews, specifically the Bank Jago application, using the Bidirectional Long Short-Term Memory (Bi-LSTM) method combined with Global Vectors for Word Representation (GloVe) Word Embedding and Synthetic Minority Over-sampling Technique (SMOTE). A total of 5,590 reviews were collected through web scraping from February 1 to September 29, 2023. Text preprocessing included cleaning, case folding, tokenization, normalization, stopword removal, and stemming. SMOTE was applied to balance the training data, resulting in a 50:50 class distribution. Hyperparameter tuning using Grid Search identified the optimal configuration of 32 hidden units and 0.0 dropout rate. The model achieved 89.09% validation accuracy and an average validation accuracy of 89.13% using 5-Fold Cross Validation. Evaluation on the test set produced an accuracy of 88.55%, demonstrating that the proposed approach effectively classifies sentiment in digital banking reviews.