Integrated Feature-Temporal GAN for Imbalanced Transaction Fraud Detection

Transaction fraud detection (TFD) poses a significant challenge due to the severe class imbalance, where fraudulent transactions, though rare, cause substantial financial losses. Existing methods often fail to adequately capture both the critical discriminative features of fraud and the temporal dependencies inherent in transaction sequences, leading to suboptimal detection performance. In this work, we introduce the Integrated Feature-Temporal-aware Generative Adversarial Network (IFT-GAN), a novel framework that simultaneously refines feature space and models temporal dynamics during the generation of synthetic samples. IFT-GAN introduces two key innovations: 1) A Feature-Aware Sample Selection (FASS) module that employs Fisher score-based feature ranking and weighted k-means clustering to create a balanced subset enriched with discriminative fraud signatures; and 2) A Temporal Dependency GAN (TD-GAN) architecture employing a novel Time-aware Gated Recurrent Unit (TaGRU) in both the generator and discriminator, explicitly modeling the temporal intervals between transactions through adaptive temporal gating. Distinct from existing methods, IFT-GAN establishes a unified framework where feature-space refinement and temporal sequence generation mutually reinforce each other. The FASS module effectively reduces the amplification of non-discriminative features, while the TD-GAN ensures temporal consistency through adversarial training. Extensive experiments on both real-world and publicly available datasets demonstrate that IFT-GAN outperforms its competitive peers across key detection metrics, indicating the effectiveness of IFT-GAN in addressing the class imbalance challenge while preserving the integrity of authentic fraud patterns.