A network with attributes associated with its nodes is called an attribute network. Detecting anomalous nodes in attribute networks has wide-ranging applications in the real world, such as identifying fake information spreaders on social media, special users causing traffic congestion in transportation networks, financial fraudsters, and telecom fraudsters in communication networks. Existing methods primarily use graph neural networks (GNNs) to learn node representations for anomaly detection. However, GNN-based detection methods suffer from issues such as inaccurate node feature learning and model overfitting. This paper proposes a novel anomaly detection model, NEAnomSCCA, based on network-enhanced joint sparse canonical correlation analysis. The model constructs an enhanced network to eliminate isolated nodes and subgraphs, ensuring the model accurately learns node features. It also incorporates random features to improve the model"s generalization ability. Building upon the GNN model for extracting node features, it calculates reconstruction errors to detect the majority of anomalies in the network. It also calculates typical sparse vectors to handle high-dimensional sparse data and detect specific anomalies. Finally, by combining the reconstruction errors and the correlation of the typical sparse vectors for each node, the anomaly score is computed to detect anomalous nodes. Comparative experiments on four real-world datasets—BlogCatalog, Cora, Citeseer, and Pubmed—were conducted against models such as JAANE, ARISE, and PROPOSED. The results show that NEAnomSCCA improves the AUC by 1.13%, 1.35%, 1.11%, and 3.33%, respectively, compared to the current best-performing models.