Publication Details
Abstract
The fast growth of digital payment systems in the United States has made the financial world turn a considerable turn and made the transactions fast, convenient, and frictionless. This rapid growth has however exposed people to more fraudulent acts like account takeovers, card-not-present fraud, spoofing of identity and unauthorized digital transactions. Conventional methods of fraud detection, specifically rule-based methods, are usually not able to adjust to changing frauds and can produce massive false-positive values, which lead to poor operational inefficiency and customer dissatisfaction. To overcome them, this study introduces an AI-based predictive modeling system that would identify and avert financial fraud in the U.S. digital payment economies. This research employs the IEEE-CIS Fraud Detection Dataset which is one of the most sophisticated and realistic publicly available dataset that has a strong reflection of behavioral, transactional, and identity-based aspects of U.S. online payments. The approach will combine a sequence of machine-learning steps based on data preprocessing, feature engineering, imbalanced-data processing strategies and sophisticated classification methods. Several machine-learning models were tested with the assistance of key performance indicators, including AUC-PR, precision, recall, and false-positive rate, and they are Logistic Regression, Random Forest, XGBoost, and LightGBM. Among them, the LightGBM model had better prediction capacity because it is able to represent the well-defined interactions among high-dimensional features with the capability of handling missing data and unbalanced classes. The techniques of explainability, especially SHAP (Shapley Additive Explanations), were used in order to justify the model decisions and determine the most effective predictors of fraudulent behavior. The results showed that the device information anomalies, inconsistent identity attributes, abrupt spending variations, and abnormal transaction schedules were some of the best indicators of fraud. These findings affirm that AI-based models have a high ability to outperform conventional techniques in detecting fraud and to provide powerful early warning systems and aiding real-time risk reduction in digital financial systems. On the whole, this study can advance the creation of intelligent, interpretable, and scalable fraud detection models that could be incorporated into the banking and fintech business in the United States to promote the financial security aspect and minimize losses related to fraud.