Enhancing Customer Contact Efficiency with Graph Neural Networks in Credit Card Fraud Detection Workflow

Credit card fraud has been a persistent issue since the last century, causing significant financial losses to the industry. The most effective way to prevent fraud is by contacting customers to verify suspicious transactions. However, while these systems are designed to detect fraudulent activity, they often mistakenly flag legitimate transactions, leading to unnecessary declines that disrupt the user experience and erode customer trust. Frequent false positives can frustrate customers, resulting in dissatisfaction, increased complaints, and a diminished sense of security. To address these limitations, we propose a fraud detection framework incorporating Relational Graph Convolutional Networks (RGCN) to enhance the accuracy and efficiency of identifying fraudulent transactions. By leveraging the relational structure of transaction data, our model reduces the need for direct customer confirmation while maintaining high detection performance. Our experiments are conducted using the IBM credit card transaction dataset to evaluate the effectiveness of this approach.

💡 Research Summary

The paper addresses a critical pain point in credit‑card fraud prevention: the excessive reliance on customer contact to verify suspicious transactions, which generates false positives, erodes trust, and inflates operational costs. To mitigate this, the authors propose a fraud‑detection framework built around Relational Graph Convolutional Networks (RGCN), a variant of Graph Neural Networks designed for heterogeneous graphs with multiple edge types.

Using the publicly available IBM credit‑card transaction dataset—comprising 24 million synthetic transactions, 100 000 unique cards, 6 000 merchants, and only 30 000 fraud cases (0.1 % prevalence)—the authors construct an undirected bipartite graph. Nodes represent cards (combined with user identifiers) and merchants, while edges encode individual transactions enriched with attributes such as amount, timestamp, MCC code, and geographic information. This graph captures the relational structure that traditional tabular models ignore (e.g., “card‑merchant”, “card‑time‑slot”, “merchant‑region” relationships).

The RGCN model processes this graph by aggregating neighbor information separately for each relation type. The update rule (Equation 2) introduces a relation‑specific weight matrix (W^{(l)}r) and normalization constant (c{ij}^r) to prevent gradient explosion, followed by a non‑linear ReLU activation. By learning distinct transformations for each edge type, the network can differentiate patterns such as a card repeatedly used at a high‑risk merchant versus occasional legitimate purchases.

Given the severe class imbalance, the authors replace the standard cross‑entropy loss with focal loss, which down‑weights easy, majority‑class examples and focuses learning on the scarce fraudulent instances. Training employs the Adam optimizer with default momentum and adaptive learning rates. The dataset is split 80 %/20 % for training and testing; early experiments show a rapid decline in both training and validation loss within the first 30 epochs, stabilizing near zero thereafter. The final test accuracy reaches 99.88 %, substantially higher than typical baselines (Random Forest, RNN, or rule‑based systems) that usually hover around 90‑95 % on similar tasks.

Beyond raw performance, the paper emphasizes operational impact: a more accurate classifier reduces the number of transactions flagged for manual verification, thereby cutting down on unnecessary customer outreach. This directly translates to lower complaint volumes, improved user experience, and reduced costs for fraud‑operations teams.

The authors also discuss limitations. While RGCN excels at modeling relational data, it incurs higher memory and computational overhead, especially when scaling to billions of daily transactions. The current study uses a synthetic dataset, so real‑world generalization remains an open question. Moreover, the paper lacks a detailed latency analysis, which is crucial for real‑time fraud decision pipelines.

In conclusion, the work demonstrates that incorporating relational graph neural networks into credit‑card fraud detection pipelines can dramatically improve detection accuracy and minimize disruptive customer contacts. Future directions include deploying incremental graph updates for streaming data, testing on proprietary bank datasets, and exploring model compression techniques (e.g., quantization, pruning) to meet production‑grade latency and resource constraints.