Researchers have introduced MaRENet, a Masked Attention and Recurrent Encoding Network designed to address the persistent difficulties of transaction-level credit card fraud detection. Identifying fraudulent transactions remains inherently difficult because unauthorized activities occur rarely, remain deeply embedded within extensive streams of legitimate consumer activity, and depend on subtle temporal cues. To resolve these operational hurdles, the proposed architecture models transaction streams sequentially while prioritizing informative behavioral signals. In comparative experimental assessments across two large-scale public transaction datasets, the new framework outperformed all evaluated baseline models, recording high F1-scores of 0.9694 on Dataset 1 and 0.9883 on Dataset 2.
The core design of MaRENet processes each transaction within the context of a fixed-length history window rather than treating transactions as isolated occurrences. Within this sequential window, the system strengthens temporal dynamics using Temporal Feature Scaling, which explicitly amplifies hour-of-day and day-of-week indicators. This structured temporal representation ensures that the contextual timeline surrounding each transaction is systematically captured before further feature extraction takes place.
Following temporal scaling, the model incorporates Channel-wise Feature Attention to calibrate the significance of individual data attributes. This attention mechanism assigns higher numerical weights to more informative transaction characteristics, effectively highlighting relevant features while dampening less diagnostic input dimensions. Consequently, the downstream recurrent encoder is directed to focus its computational capacity on the specific features most instrumental for accurate fraud detection. The entire architecture is trained using binary cross-entropy, optimizing the network to distinguish reliably between rare fraudulent actions and common legitimate purchases across diverse payment records.
To evaluate real-world utility, the researchers performed additional validation using a time-ordered data split, training MaRENet strictly on earlier transactions and testing it on later transactions. These evaluations confirmed that the model remains robust and effective when encountering future transactional sequences over time. Furthermore, in hardware efficiency assessments under a reported GPU inference benchmark, the MaRENet-GRU variant demonstrated lower operational latency and higher processing throughput than the comparative TH-LSTM model, while simultaneously maintaining fewer trainable parameters.
Subject of Research: Deep learning architecture for transaction-level credit card fraud detection
Article Title: MaRENet: Masked attention and Recurrent Encoding for transaction-level credit card fraud detection
Article References: Nguyen, D.-A., Thi Nguyen, K.-N., Nguyen, T.-N., Ngo, V.-D., Vu, H., Nguyen, T.-C., Xuan Trang, M., & Van Luong, T. (2027). MaRENet: Masked attention and Recurrent Encoding for transaction-level credit card fraud detection. Cluster Computing, 30(1), Article 1. https://doi.org/10.1007/s10586-026-06747-w
Image Credits: AI Generated
DOI: 10.1007/s10586-026-06747-w
Keywords: credit card fraud detection, MaRENet, masked attention, recurrent encoding, transaction-level detection, temporal feature scaling, channel-wise feature attention, binary cross-entropy, recurrent encoder, fixed-length history window, time-ordered split, GPU inference benchmark, MaRENet-GRU, TH-LSTM, throughput, latency, F1-score
News Source: Denise Maddox. (November 16, 2026). MaRENet Outperforms Baselines in Transaction-Level Credit Card Fraud Detection. Scienmag.



