Effective High-order Graph Representation Learning for Credit Card Fraud Detection

TL;DR

The paper tackles credit card fraud detection under disguise by leveraging high-order, pure representations learned from decoupled multi-hop transaction graphs. It introduces HOGRL, a framework that constructs high-order graphs with $S^{l}=A^{l}-A^{l-1}+I$, learns order-specific embeddings via a mixture-of-experts, and combines them with original-graph embeddings to preserve multi-hop dependencies. An MLP detector operating on the final representations achieves fraud probabilities, optimized with cross-entropy and Adam. Experiments on real and public datasets show that HOGRL consistently outperforms state-of-the-art baselines in AUC, F1-macro, and GMean, with interpretability evidenced by increased fraud node homophily and clearer embedding separations. The approach offers robust, scalable improvements for high-order fraud camouflage scenarios and has practical significance for financial security systems.

Abstract

Credit card fraud imposes significant costs on both cardholders and issuing banks. Fraudsters often disguise their crimes, such as using legitimate transactions through several benign users to bypass anti-fraud detection. Existing graph neural network (GNN) models struggle with learning features of camouflaged, indirect multi-hop transactions due to their inherent over-smoothing issues in deep multi-layer aggregation, presenting a major challenge in detecting disguised relationships. Therefore, in this paper, we propose a novel High-order Graph Representation Learning model (HOGRL) to avoid incorporating excessive noise during the multi-layer aggregation process. In particular, HOGRL learns different orders of \emph{pure} representations directly from high-order transaction graphs. We realize this goal by effectively constructing high-order transaction graphs first and then learning the \emph{pure} representations of each order so that the model could identify fraudsters' multi-hop indirect transactions via multi-layer \emph{pure} feature learning. In addition, we introduce a mixture-of-expert attention mechanism to automatically determine the importance of different orders for jointly optimizing fraud detection performance. We conduct extensive experiments in both the open source and real-world datasets, the result demonstrates the significant improvements of our proposed HOGRL compared with state-of-the-art fraud detection baselines. HOGRL's superior performance also proves its effectiveness in addressing high-order fraud camouflage criminals.

Figures (6)

• Figure 1: Node homophily distribution across datasets. Node homophily indicates the proportion of neighbors with the same label.
• Figure 2: The illustration of the proposed HOGRL model architecture. It contains four components: (a) Graph construction based on original transactions. We treat each layer of the GNN network as an expert network and dynamically allocate weights using a mixture-of-expert attention mechanism. (b) Embedding generation based on multi-layer high-order graphs. (c) Graph neural network embedding generation based on the original graph. (d) Detection network based on concatenated embeddings and joint optimization.
• Figure 3: Results of models with different layers.
• Figure 4: Parameter sensitivity analysis with respect to (a) the hidden dimension; (b) the weight $\gamma$, (c) and (d) the train ratio.
• Figure 5: Homophily density distribution with different layers. The left side illustrates the homophily statistics of the traditional mixed-order propagation process, while the right side shows the homophily statistics of the high-order graphs proposed by HOGRL.