Multi-Objective Explainable Transformer Framework for Fraud Risk Forecasting in Digital Payment Ecosystems

Abstract

Background: The high rate of digital payment systems has led to rapid rate of transactions and increased vulnerability to financial fraud, thus leading to need to have intelligent, transparent, and proactive fraud monitoring solutions. Current fraud detection tools conceptualize the issue as binary classification, and act as black-box predictors, constraining their ability to provide future risk awareness and interpretability needed to comply with requirements in regulation and operational decision-making. Methods: The paper presents an explainable, multi-objective Transformer model to predict fraud in urban digital payment ecosystems. It optimises real-time fraud detection and constantly-in-the-future risk estimation using a Transformer Temporal encoder and dual-output architecture for classification and regression. Explainable artificial intelligence and SHAP-based temporal attention visualisation enhance transparency and identify significant transaction attributes influencing predictions. Results: Large-scale experimental results on a large-scale dataset of digital payment transactions show that the multi objective model can be effective in terms of competitive fraud detection with ROC-AUC of 0.843 and at the same time predict future risk of fraud with an RMSE of 0.049. Competitive comparison with classical machine learning classifiers and deep sequential baselines can prove the usefulness of the proposed strategy in balancing the predictive and forecasting performances as well as their interpretability. Conclusion and Implication: The proposed framework can be viewed as a clear and active way of dealing with fraud monitoring in modern digital payment systems, as it combines the power of multi-objective learning and explainable modelling. The suggested framework optimises both, complementary, goals in real-time fraud detection and constantly-in-the-future risk of fraud estimation. Novelty: In contrast to the many single-task fraud detection models, this research demonstrates a joint multi-objective Transformer model for fraud classification and future risk forecasting, along with complementary explainability features via SHAP-based feature attribution and temporal attention analysis, which translates to transparent and anticipatory fraud risk management in digital payment ecosystems.