Anomaly Detection in Financial Payment Transactions Using Efficient Data-Driven Machine Learning Techniques

Sophisticated fraud cases due to the growth of Digital Payment Systems occurred within an environment where thousands of new forms of Digital Payments were developed every month; thus, there must be an increasing demand for effective reasons for detecting Financial Transaction anomalies. This research will focus on developing a framework using ML to detect fraudulent payment behaviors within a very imbalanced Credit Card dataset. The proposed approach will incorporate several key processes of data mining including Exploratory Data Analysis, Duplicates Removal, Robust Scaling (through Robust Scaler), SMOTE (Synthetic Minority Over Sampling Technique), and Three Models (XGBoost, Random Forest and Hybrid XGBoost with Isolation Forest anomaly scoring feature). From the results of this experiment, it shows that the proposed Hybrid XGBoost model had the highest Accuracy (99.95%), Precision (97.22%) and F1-Score (83.83%) of all models tested, thus outperforming all base models. The originality of the study lies in the integration of Unsupervised Anomaly Scores with a Supervised Classifier, which has been used to produce a system that is more sensitive to the detection of slight fraudulent patterns and at the same time reduces the rate of false positives. This validates the effectiveness and scalability of the proposed model to practical Financial Fraud Detection in the Financial Industry.