Joint Modelling of Transaction Intent and Fraud Risk Using TransformerBased Embeddings in Real-Time Payment Systems

Abstract

The rapid expansion of digital payments has significantly increased the need for effective fraud detection systems to ensure secure financial transactions. Traditional methods, such as manual verification, rule-based systems, and basic statistical techniques, were time-consuming, labourintensive, and often failed to identify complex fraud patterns in large-scale data. These limitations made them unsuitable for real-time decision-making in modern financial environments. To overcome these challenges, this study proposes an intelligent web-based system for fraud detection in Unified Payments Interface (UPI) transactions using advanced Machine Learning (ML) and Deep Learning (DL) techniques. The system is developed using the Flask (Web Framework) and integrates Natural Language Processing (NLP) to analyse textual transaction data. It employs Sentence-Bidirectional Encoder Representations from Transformers (SBERT) to generate contextual embeddings that capture meaningful patterns within transaction details. These embeddings are used to train multiple classification models, including Gaussian Naive Bayes (GNB), Bernoulli Naive Bayes (BNB), Multinomial Naive Bayes (MNB), and the proposed Histogram-based Gradient Boosting (HGB) classifier. The system performs both binary classification for fraud detection and multi-class classification for identifying transaction types. Experimental results indicate that the HGB model achieves superior performance in terms of accuracy, precision, recall, and F1-score compared to other models. Furthermore, the system features a secure login interface that enables users to upload datasets, generate real-time predictions, and download results. This solution enhances scalability, automation, and accuracy, making it highly effective for modern digital fraud detection