Abstract
Anomaly detection in attributed networks has attracted considerable interest owing to its applications in security, fraud detection, and system monitoring. These networks pose unique challenges as anomalies may arise from either structural irregularities or inconsistencies in node attributes, or both. Existing approaches typically focus on a single learning paradigm, such as reconstruction-based modeling or contrastive representation learning, which limits their effectiveness in capturing the full spectrum of anomalous behaviors. In this work, we propose a comprehensive hybrid framework that jointly integrates four learning components: graph structure reconstruction, attribute reconstruction, community-aware contrastive learning, and similarity-aware anomaly scoring. This unified design allows the model to learn both global structural patterns and local semantic consistencies, while also promoting discriminative representation learning and neighborhood-based anomaly refinement. We evaluate the proposed framework on six widely adopted benchmark datasets- BlogCatalog, Flickr, ACM, Cora, Citeseer, and Pubmed. Extensive experiments demonstrate that our method significantly outperforms state-of-the-art baselines across two evaluation metrics: AUC and AUPR. The ablation study confirms the individual contributions of each module, and the parameter sensitivity analysis reveals that the framework remains robust under varying hyperparameter conditions. These results validate the effectiveness and generalizability of the proposed framework in detecting a wide range of anomalies in complex attributed graphs.