Sentiment Analysis of BYOND by BSI User Reviews Using TF-IDF Feature Extraction and XGBoost Classification

Abstract

Abstract. The migration of Bank Syariah Indonesia (BSI) from BSI Mobile to the BYOND by BSI super-app led to a sharp increase in Google Play Store reviews, generating high-dimensional, unstructured text data. Single classifiers, such as Decision Tree, are prone to overfitting and unstable predictions because their performance depends heavily on hyperparameter settings. This study applies Extreme Gradient Boosting (XGBoost), an ensemble learning method, with Term Frequency–Inverse Document Frequency (TF-IDF) feature extraction to classify user sentiment toward BYOND by BSI into positive and negative classes. The dataset comprises 51,475 Indonesian-language reviews collected between October 2024 and December 2025. Ratings of 1–3 were labeled negative and 4–5 positive. After preprocessing, 51,232 reviews remained, with a balanced class distribution of 52.0% negative and 48.0% positive. TF-IDF generated a 3,000 feature matrix with 99.66% sparsity. The dataset was split into 80% training and 20% testing sets, and hyperparameters were optimized using Randomized Search with 3-fold cross-validation. The optimized XGBoost model achieved an F1-score of 92.28% on the testing set, with the word easy (“mudah”) identified as the most influential feature. These findings provide BSI with a reliable, data-driven basis for improving its digital services.