A Comparative Performance Analysis of Naïve Bayes, LSTM, and BiLSTM with Data Balancing Techniques for Sentiment Analysis of EasyCash Application Reviews

Abstract

This study compares the performance of Naïve Bayes, Long Short-Term Memory (LSTM), and Bidirectional LSTM (BiLSTM) models in sentiment analysis of EasyCash application reviews, with data balancing techniques applied throughout the process. The dataset was collected from the Google Play Store and processed through cleaning, tokenization, stemming, and normalization. Sentiment labeling classified reviews into positive, neutral, and negative categories. To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied prior to model training. Feature extraction was conducted using TF-IDF, and models were evaluated on accuracy, precision, recall, and F1-score. Naïve Bayes outperformed both LSTM and BiLSTM, producing higher accuracy and more stable results across evaluation metrics. The findings suggest that simpler machine learning models can be more effective than deep learning approaches when working with limited and imbalanced datasets. Careful data preprocessing, appropriate balancing techniques, and deliberate model selection remain central to achieving reliable sentiment classification performance in fintech applications.