Are There Any Hidden Agents in Your Recommendations? Anomaly Detection via Structure Purification and Stability Verification

The reliability of training data has become a crucial concern for recommender systems, as real-world interaction logs may contain long-tail noise or even hidden manipulation. To address these challenges, we propose HEAR, an anomaly detection framework that evaluates recommendation security from the perspective of data stability. HEAR follows a “discovery-first, detection-later” paradigm. First, we encode user ratings into a co-occurrence association graph and remove long-tail structural noise via multi-round tight community mining. Second, we introduce a sparse Rayleigh-quotient-driven suspicious subgraph extraction module to isolate a max-suspicious subgraph that contains anomaly-dense regions. On this purified structure, we design a Lyapunov stability diagnosis based on embedding deviation to determine whether the dataset has a systemic poisoning risk without labels. Finally, we adopt orthogonal bi-hypersphere boundary learning in the embedding space to separate “suspicious-benign-suspicious” regions and identify malicious users with low false alarm rates (FARs). Compared with competitive baselines, HEAR achieves an average improvement of 14.0% in terms of detection rate and a 0.32% reduction in the FAR. Furthermore, HEAR identifies potential malicious review patterns in real-world unlabeled data, demonstrating its effectiveness in practical scenarios.