Sentiment Analysis of AdaKami Online Lending Application Reviews: Replication Data and Code

Replication package for a study comparing Naive Bayes Classifier (NBC), Support Vector Machine (SVM), and Random Forest (RF) for three-class sentiment classification of Indonesian peer-to-peer lending application reviews. The dataset contains 35,499 Indonesian-language user reviews of the AdaKami application, collected from the Google Play Store for the period January to December 2025. After preprocessing, 32,905 reviews entered the modelling stage. Sentiment labels were derived from star ratings and validated against manual annotation by two independent annotators, yielding a Cohen's kappa of 0.8595. The package includes the raw and preprocessed corpora, the manual annotation sample, six Jupyter notebooks covering the full CRISP-DM pipeline from scraping to evaluation, the three trained classifiers, the fitted TF-IDF vectoriser, and the source code of the accompanying Streamlit dashboard. Naive Bayes achieved the highest macro F1-score (0.6416), followed by SVM (0.6299) and Random Forest (0.6134), against a majority-class baseline of 0.2623. A Wilcoxon signed-rank test on per-fold scores returned W = 0, p = 0.0625, so the difference between NBC and SVM is not statistically significant at alpha = 0.05. Privacy: the reviewer username column captured during scraping was never used in any analysis step and has been removed from every file in this deposit. All reviews were screened for phone numbers, email addresses, national identity numbers, and bank account numbers, with no matches found. See README.md for the full data statement and the removal-request procedure.