Artificial Intelligence for Imbalanced Financial Classification and Its Societal Implications for Digital Trust: A Cost-Sensitive XGBoost–Transformer Ensemble

Abstract

Artificial intelligence (AI) increasingly mediates how financial institutions decide who is trusted, who is scrutinised, and who is refused — decisions that carry direct social consequences for individuals and communities. Automated credit-risk and fraud-assessment systems are deployed at a scale where their errors are not merely statistical but socially consequential, because they can restrict access to credit, widen financial exclusion, and erode public confidence in digital finance. This paper addresses one technical precondition for trustworthy deployment: reliable learning from severely imbalanced financial data. It treats the task as what the data actually measure — prediction of credit-card default rather than verified fraud — and proposes FraudShield, a hybrid framework that combines a gradient-boosting model (XGBoost), a Transformer-based tabular model, an adaptive gating mechanism that learns a per-instance weighting of the two branches, and a cost-sensitive decision layer that reflects the unequal social and economic cost of different errors. The framework is evaluated on the UCI Credit Card Default dataset (30,000 clients), whose positive (default) class comprises 22.12% of the observations — a ratio of approximately 1:3.52. The evaluation protocol splits the data before any preprocessing, fits target encoding and standardization within each training fold only, and applies SMOTE to the training folds alone. Under this corrected protocol, FraudShield attains an AUC-ROC of 0.943 and an F1-score of 0.892, and the cost-sensitive layer raises measured cost savings from 16.7% to 28.4% over the majority-class baseline. The paper interprets these results within a broader discussion of AI accountability, fairness, and digital trust, and argues that technically sound imbalanced-learning methods are a necessary — though not sufficient — condition for socially responsible automated decision-making in finance.