Credit Card Fraud Detection Using RoFormer Model With Relative Distance Rotating Encoding

TL;DR

This paper presents a novel fraud detection method using the RoFormer model enhanced with Relative Distance Rotating Encoding to better capture temporal dependencies in transaction data.

Contribution

It introduces ReDRE into RoFormer, improving the model's ability to detect transactional fraud by capturing temporal relationships more effectively.

Findings

Enhanced fraud detection accuracy

Better modeling of temporal dependencies

Improved characterization of transaction sequences

Abstract

Fraud detection is one of the most important challenges that financial systems must address. Detecting fraudulent transactions is critical for payment gateway companies like Flow Payment, which process millions of transactions monthly and require robust security measures to mitigate financial risks. Increasing transaction authorization rates while reducing fraud is essential for providing a good user experience and building a sustainable business. For this reason, discovering novel and improved methods to detect fraud requires continuous research and investment for any company that wants to succeed in this industry. In this work, we introduced a novel method for detecting transactional fraud by incorporating the Relative Distance Rotating Encoding (ReDRE) in the RoFormer model. The incorporation of angle rotation using ReDRE enhances the characterization of time series data within a Transformer, leading to improved fraud detection by better capturing temporal dependencies and event relationships.