Most graph neural networks predominantly rely on local message passing, which limits their ability to capture high-order interactions between distant graph nodes and to understand the broader context of a node within the graph. This limitation often leads to missed detections of anomalous nodes, as the network fails to leverage the full structural and contextual information embedded in the graph. In this chapter, we introduce a flexible graph encoder–decoder model for unsupervised anomaly detection in attributed networks. In the encoding stage, we design a flexible graph convolutional encoder that allows for effective aggregation of information from higher-order node neighborhoods, improving the encoder’s ability to learn from the global structure of the graph while maintaining computational efficiency. The proposed encoder provides theoretical guarantees of stability due to its propagation matrix design, ensuring numerically stable feature updates. In the decoding stage, a structure decoder predicts the presence or absence of edges between nodes, while an attribute decoder reconstructs node features using the graph structure and latent representations. We conduct comprehensive empirical evaluations of our encoder–decoder model on six benchmark datasets using several evaluation metrics. The results demonstrate the superior performance of our model over competing anomaly detection approaches, highlighting its effectiveness in identifying anomalous nodes in attributed networks.