<p>Anomaly detection in attributed networks is a critical task with applications in cybersecurity, social network analysis, and fraud detection. However, existing methods often suffer from over-smoothing, limited contextual representation, and inadequate fusion of structural and attribute information. This paper proposes a novel framework that integrates multi-level feature aggregation (MLFA), deep skip-gram embedding, and multi-head attention to effectively detect anomalies in attributed graphs. The MLFA module captures node-, neighborhood-, and community-level features using a combination of Weisfeiler–Lehman aggregation and graph attention networks. These multi-scale features are encoded through a deep skip-gram model to learn context-aware embeddings, which are further refined via a multi-head attention mechanism to capture diverse relational dependencies. The final anomaly scores are computed by jointly considering feature reconstruction loss and contextual deviation, ensuring robustness to both attribute noise and topological irregularities. Extensive experiments conducted on real-world datasets, including BlogCatalog and Flickr, demonstrate that the proposed method outperforms state-of-the-art baselines such as DOMINANT, ARISE, and DVAEGMM, achieving superior AUC and average precision scores. The results validate the effectiveness of hierarchical representation and attention-based fusion for anomaly detection in complex attributed networks.</p>