Muhammad Bayu Wijaya Putra, Rinto Priambodo

Graph-Based Hybrid GNN-Transformer for Imbalanced Credit Card Fraud Detection

Introduction

Graph-based hybrid gnn-transformer for imbalanced credit card fraud detection. Detect imbalanced credit card fraud with a novel Graph-based Hybrid GNN-Transformer. It combines relational learning and feature interaction for improved detection sensitivity.

Abstract

Credit card fraud detection faces two major challenges: severe class imbalance and the limited ability of conventional feature-based models to capture relational patterns among transactions. This study proposes a graph-based Hybrid GNN-Transformer architecture for imbalanced credit card fraud detection by integrating transaction-level relational learning through k-nearest neighbor graph construction and feature-interaction learning through multi-head self-attention. The novelty of this study lies in combining graph-based transaction modeling and Transformer-based feature interaction within a unified architecture. Using the selected graph configuration and validation-based threshold tuning, the proposed model achieved 79.71% precision, 74.32% recall, 76.92% F1-score, 96.06% ROC-AUC, and 68.65% PR-AUC. Compared with Logistic Regression, Random Forest, and Gradient Boosting baselines, the hybrid model showed competitive fraud detection sensitivity, although the baseline classifiers still achieved stronger overall F1-score and PR-AUC. Ablation results show that the hybrid architecture improves minority-class detection compared with single-branch variants by combining relational transaction information from the GNN branch and feature-interaction information from the Transformer branch. These findings indicate that graph-based hybrid representation learning is a promising direction for imbalanced fraud detection, while further optimization is still required to improve precision-recall balance and competitiveness against strong feature-based baselines.

Review

This study introduces a Graph-based Hybrid GNN-Transformer architecture to tackle the dual challenges of severe class imbalance and the inadequacy of traditional models in capturing relational patterns in credit card fraud detection. The core innovation lies in its unified approach, integrating a GNN branch for transaction-level relational learning via k-nearest neighbor graph construction with a Transformer branch for feature-interaction learning through multi-head self-attention. This conceptual design elegantly addresses a critical gap in existing methodologies by simultaneously modeling explicit relationships between transactions and intricate feature dependencies within them, offering a novel hybrid representation learning paradigm for a high-stakes detection problem. The proposed hybrid model demonstrates commendable performance, particularly in terms of ROC-AUC at 96.06%, indicating strong discriminative power between classes. It achieves a precision of 79.71%, recall of 74.32%, and an F1-score of 76.92%, with a PR-AUC of 68.65%. While the abstract notes competitive fraud detection sensitivity compared to baselines like Logistic Regression, Random Forest, and Gradient Boosting, it critically acknowledges that these conventional classifiers surprisingly still yielded stronger overall F1-score and PR-AUC. The ablation study, however, provides crucial insight, affirming that the hybrid architecture indeed improves minority-class detection by synergistically combining relational information from the GNN branch and feature-interaction insights from the Transformer branch, thus validating the architectural rationale. Overall, the paper presents a promising direction for imbalanced fraud detection through its innovative hybrid GNN-Transformer architecture. The identified improvement in minority-class detection through the combined approach is a significant finding. However, the abstract's candid admission regarding the baseline models' superior overall F1-score and PR-AUC suggests that while the architectural integration is sound, further optimization is imperative. Future work should focus on refining the model's ability to achieve a better precision-recall balance and enhance its competitiveness against strong feature-based baselines, perhaps by exploring advanced graph construction techniques, alternative attention mechanisms, or more sophisticated loss functions tailored for extreme imbalance, to fully realize the potential of this compelling hybrid framework.

Full Text

Comments

You need to be logged in to post a comment.