Abstract: GNNs have emerged as a dominant approach for fraud detection due to their inherent ability to model graph-structured data and capture complex relational patterns in fraudulent activities.However,existing GNN-based fraud detection models face critical limitations:homogeneous GNNs struggle with heterogeneous relationships in fraud graphs,while heterogeneous GNNs typically process such relationships in only a single attribute or structural space,restricting their detection performance.To overcome these challenges,this paper proposes a novel dual-space heterogeneous GNN for fraud detection,which models user relationships as a multi-relational heterogeneous directed graph and employs a multi-layer graph convolutional architecture.Each convolutional layer integrates three key modules:(1)a heterophily learning module that separately learns node heterophily in attribute and structural spaces using labeled node information,then fuses these features via a weighted strategy;(2)a cross-space graph aggregation module that computes attention weights from the fused heterophily and updates node representations through multi-relational aggregation;(3)a prototype-guided classification module that constructs class prototypes from labeled nodes to guide the classification of unlabeled nodes.To address data scarcity and class imbalance,the model adopts a balanced sampling strategy for semi-supervised training.Experimental results on YelpChi and Amazon datasetss demonstrate that the proposed model has signi-ficant improvements,with Recall increasing by 0.962 6% and 0.644 4%,and AUC rising by 0.859 4% and 0.147 9%,respectively,outperforming nine baseline models.

Key words: Fraud detection, Graph neural networks, Heterophily learning, Prototype learning, Multi-relational graphs