Deteksi Ulasan Palsu pada Kafe di Google Maps Menggunakan Stacking SVM dan XGBoost Berbasis Teks dan Perilaku Pengguna

Abstract

Fake reviews on Google Maps pose a serious threat to the credibility of digital information, particularly in the food and beverage sector where online reputation is critical. The rapid growth of cafes in Jember Regency underscores the need for an automated detection system to protect consumers and promote fair business competition. This study aimed to develop a fake review detection system using a machine learning-based stacking approach that integrates textual features and user behavioral features. Data were collected through web scraping from Google Maps, resulting in 16,978 validated reviews. Labeling was performed using Latent Dirichlet Allocation (LDA) with an optimal topic count of k=4 (Coherence Score: 0.6487), yielding 5,108 genuine and 11,870 fake reviews. A two-level stacking architecture was constructed using SVM with TF-IDF representation and XGBoost with heuristic features as Level-0 base-learners, and XGBoost as the Level-1 metalearner. Class imbalance was addressed using BorderlineSMOTE on the training data. Evaluation results showed that the stacking model achieved the highest performance with an F1-Score of 0.9177 and ROC-AUC of 0.9569, surpassing standalone SVM (F1-Score: 0.8944, AUC: 0.9318) and XGBoost (F1-Score: 0.8486, AUC: 0.9149). Pattern analysis revealed that fake reviews tend to use generic and promotional language, excessive emojis, and extreme ratings, while genuine reviews are more descriptive and contextual. The system was implemented as a web application featuring automatic review detection and rule-based cafe recommendations. This study demonstrates the effectiveness of combining textual and behavioral features within a stacking architecture for detecting fake reviews in the Indonesian local context.