FraudTransformer: Time-Aware GPT for Transaction Fraud Detection

2025-09-28

AI Summary

Detecting payment fraud in real-time banking transaction streams poses significant challenges due to the irregular temporal dynamics and complex sequential dependencies inherent in such data. Method: This paper proposes a sequence-based approach that jointly models temporal order and irregular inter-event intervals. Its core innovation lies in a dedicated time encoder and a learnable positional encoder, which collaboratively capture absolute timestamps and relative event intervals to enhance sensitivity to dynamic temporal patterns. Built upon the GPT architecture, the model is trained end-to-end on large-scale, industrial-grade transaction sequences. Contribution/Results: Experiments on real-world banking data demonstrate substantial improvements over strong baselines—including logistic regression, XGBoost, and LightGBM—achieving state-of-the-art AUROC and PRAUC scores. These results empirically validate the effectiveness of explicit temporal modeling for identifying sophisticated, time-sensitive fraud patterns.

Technology Category

• Machine Learning: Time-Series/Data Streams
• Data Mining & Knowledge Management: Mining of Spatial, Temporal or Spatio-Temporal Data
• Planning, Routing, and Scheduling: Temporal Planning

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs

Abstract

Detecting payment fraud in real-world banking streams requires models that can exploit both the order of events and the irregular time gaps between them. We introduce FraudTransformer, a sequence model that augments a vanilla GPT-style architecture with (i) a dedicated time encoder that embeds either absolute timestamps or inter-event values, and (ii) a learned positional encoder that preserves relative order. Experiments on a large industrial dataset -- tens of millions of transactions and auxiliary events -- show that FraudTransformer surpasses four strong classical baselines (Logistic Regression, XGBoost and LightGBM) as well as transformer ablations that omit either the time or positional component. On the held-out test set it delivers the highest AUROC and PRAUC.

Problem

• Detecting payment fraud in real-time banking transaction streams
• Modeling transaction sequences with irregular time gaps
• Improving fraud detection accuracy over classical machine learning methods

Innovation

• GPT architecture enhanced with time encoding
• Learned positional encoder for event sequence
• Combines time gaps and event order