PENGEMBANGAN KINERJA MODEL DETEKSI ULASAN PALSU MENGGUNAKAN DISTILBERT PADA PLATFORM E-COMMERCE

Various aspects of life have become easier thanks to technological advances, including business and economics. The development of digital businesses, particularly e-commerce, enables effective and efficient interactions between sellers and buyers. Customer reviews can be a primary source of information about product and service quality during the purchasing decision-making process. Therefore, the credibility and authenticity of reviews are crucial for maintaining user trust and platform integrity.
The dataset used in this study consists of 887 Indonesian-language reviews collected from one of the stores on the Tokopedia platform. A Natural Language Processing (NLP)-based approach using the DistilBERT model was used to build a fake review detection system. Cleaning, case folding and tokenization were part of the preprocessing process. The model was trained with scenarios with and without oversampling. For evaluation, accuracy, precision, recall, F1-score, confusion matrix and ROC-AUC metrics were used.
With an accuracy of 97%, an F1-score of 0.97, and a ROC-AUC of 0.9921, the model with oversampling at 50 epochs performed best. Furthermore, the high ROC-AUC value indicates the model's ability to differentiate between genuine and fake reviews. The final system was implemented into a Streamlit-based web application that can detect fake reviews and assist users in making better decisions on e-commerce platforms.
Keywords: Fake Review, DistilBERT, Natural Language Processing (NLP), Streamlit