<![CDATA[The increasing use of paylater services in various regions of Indonesia has given rise to a large collection of user reviews that contain meaningful information about how users evaluate and perceive the experience of using the service. This research focuses on the classification of sentiment contained in paylater app reviews using two machine learning approaches, namely Random Forest and Logistic Regression, which are tested both with and without the application of the Synthetic Minority Oversampling Technique (SMOTE) to address class inequality. Review information is obtained from the Google Play Store and goes through a series of initial steps involving text cleaning, case processing, standardization of non-standard words, splitting sentences into tokens, removing meaningless words, and also stemming. Subsequent characteristic extraction is carried out using the TF-IDF method (Term Frequency-Inverse Document Frequency), before the data is divided into training and testing sets using three split configurations: 80:20, 70:30, and 60:40, to evaluate model consistency across varying training data sizes. The 80:20 split consistently produced the highest performance across all models. The results of all tested configurations, the combination of Logistic Regression with SMOTE provided the best results, achieving 88.38% accuracy, 88.37% precision, 88.38% recall, and 88.37% F1-score. Unlike previous studies that analyzed sentiment from a single paylater application using a single algorithm without class balancing, this study contributes by simultaneously collecting data from three major paylater applications and empirically comparing the effect of SMOTE on two algorithms across three data split configurations, providing a more comprehensive and generalizable benchmark for paylater sentiment classification in Indonesia. This finding indicates that the application of SMOTE also strengthens the model's performance by addressing the imbalance between sentiment classes, while Logistic Regression is proven to be able to recognize patterns related to sentiment in review text. In general, this study shows that combination of TF-IDF, Logistic Regression, and SMOTE builds an effective system for classifying sentiment in paylater application reviews.]]>
