An Enhanced Fraud Detection for Recommendation Systems

Recommendation systems are increasingly targeted by fraudulent activities, also known as shilling attacks, where malicious actors manipulate ratings to distort recommendations, leading to compromised trust and system integrity. While literature offers advanced AI models such as deep learning and graph-based methods, these often face scalability and resource constraints in real-world applications. To address this, we adopt a lightweight and interpretable solution, prioritizing practicality over architectural complexity. This research introduces a novel approach by applying graph-based feature engineering to extract new features that capture relational dynamics and the statistical properties of connected nodes in recommendation systems. These features enrich the dataset by modeling relational and connectivity patterns, enabling a deeper understanding of fraudulent behaviors. Additionally, data imbalance is addressed using SMOTE algorithm, and Gradient Boosting models are trained to classify fraudulent activities. Using the YelpCHI dataset, the proposed method achieved an F1-macro score of 0.914 and GMean of 0.883, demonstrating high accuracy and robustness. Analyses results highlight the potential of combining advanced feature engineering with efficient m achine learning techniques for scalable fraud detection in recommendation systems.