Graph anomaly detection is widely used in risk identification and fraud detection in fields such as credit cards and digital currencies, owing to the powerful ability of graph-structured data to abstract and model real-world relationships. However, when anomalous nodes have a large number of associated edges with normal nodes, GNN-based algorithms often perform poorly. This is a common challenge faced by these sensitive domains. To address the low detection rate of anomalous nodes in graph-based anomaly detection, we propose an Up-to-down Subgraph Partitioning Graph Neural Network (USP-GNN). First, we enhance node feature representations by extracting structural information from the graph using filters. Next, we propagate the features of the source and target nodes to the edge features according to the directed nature of the edges. Finally, we perform clustering based on the updated edge features to partition the graph into subgraphs. Anomaly detection is then performed on each subgraph, effectively isolating the neighbors of anomalous nodes and improving detection rates. We demonstrate the effectiveness of our approach on six real anomaly detection datasets.