Learning with Blacklists on Graphs via Prototype-Guided Dual-Frequency Filtering

Abstract

Graph-based fraud detection plays a critical role in identifying anomalous accounts and preventing financial losses in real-world systems, where graphs often contain millions of nodes but only a limited number of blacklist labels are available. Existing graph neural network approaches typically rely on full-graph message passing and sufficient supervision, which leads to weak neighborhood aggregation under scarce labels and poor scalability on large-scale graphs. In this work, we propose a scalable fraud detection framework that integrates prototype-guided graph abstraction with dualfrequency filtering to effectively learn from limited blacklists. Specifically, we construct class prototypes from the scarce labeled anomalies to capture compact class-level patterns and use them to guide subgraph abstraction and edge pruning, reducing the computational burden of graph message passing. On the resulting abstracted subgraphs, we apply dual-frequency graph filters to jointly model low-frequency homogeneous signals and high-frequency heterogeneous interactions. To further enhance robustness under incomplete supervision, we leverage high-confidence pseudo-labels to refine the graph abstraction and perform a second round of efficient dual-frequency filtering for prediction refinement. Extensive experiments on four benchmark datasets, including a large-scale graph, demonstrate that the proposed method consistently outperforms state-of-the-art approaches.