Implementasi IndoBERT dan Local Outlier Factor untuk Mendeteksi Ulasan Palsu pada E-Commerce Shopee Berbasis Analisis Tekstual dan Lonjakan Ulasan

Abstract

Product reviews on e-commerce platforms like Shopee have become a crucial factor in consumer purchasing decisions. However, this influence is exploited through the practice of fake reviews, which can mislead consumers and damage the integrity of the digital market. This research aims to develop and implement an accurate fake review detection system using a hybrid approach that combines textual content analysis and behavioral analysis. The proposed method integrates IndoBERT, a language model specifically pre-trained for the Indonesian language, to extract semantic features from review texts. In parallel, behavioral analysis is conducted using the Local Outlier Factor (LOF) algorithm to identify review burst anomalies based on metrics of review count, average rating, and rating entropy. Features from both analytical paths are then combined (feature fusion) and classified using an XGBoost model. The evaluation results show that the hybrid model achieved an accuracy of 93%, significantly outperforming the baseline text-only model (87%). Furthermore, user testing results indicate that the developed application prototype was very well-received, achieving a Task Success Rate score of 92.85. This study concludes that a hybrid approach integrating linguistic and behavioral analysis is highly effective in improving the accuracy and reliability of fake review detection systems within the Indonesian e-commerce context.