Unbiased Rectification for Sequential Recommender Systems Under Fake Orders

A team of researchers has developed a novel method called DITaR (Dual-view Identification and Targeted Rectification) to defend sequential recommender systems against the growing threat of fake orders. These malicious interactions—including click farming, irrelevant item substitutions, and sequential perturbations—are embedded within genuine user sequences to manipulate exposure rates and mislead recommendations. Unlike traditional attacks that inject fake users, fake orders directly disrupt learned user preferences, creating a more insidious form of bias that degrades system trustworthiness and performance.

DITaR's core innovation is its dual-view approach to detection and its targeted rectification strategy. The method first obtains differentiated representations of user-item interactions from both collaborative (user behavior patterns) and semantic (item/content relationships) views to precisely identify suspicious fake orders. Crucially, the researchers recognize that not all fake orders are harmful; some can have a data augmentation effect. Therefore, DITaR filters its detections to select only the truly harmful samples for rectification using gradient ascent, which surgically removes bias while preserving useful information.

This targeted approach ensures the original data volume and sequence structure are maintained, protecting the system's overall performance without the enormous computational and time costs of full model retraining. The paper reports that extensive experiments on three benchmark datasets demonstrate DITaR's superior performance compared to state-of-the-art methods, achieving significant gains in recommendation accuracy (NDCG and Recall metrics), computational efficiency, and overall system robustness against adversarial manipulation.