Unsupervised Anomaly Detection In Financial Transactions Using Machine Learning

Abstract

Financial transaction systems are increasingly vulnerable to sophisticated fraudulent and anomalous activities, necessitating intelligent and scalable detection mechanisms. Machine learning–based anomaly detection has emerged as a critical solution for identifying high-risk financial behaviors in large-scale transactional environments. This study presents an advanced anomaly detection framework utilizing the Financial Transaction and Risk Management dataset from Kaggle. The workflow incorporates comprehensive data preprocessing, feature engineering (temporal extraction, group-based aggregation, log transformation, and interaction features), encoding, ADASYN-based resampling, and Min–Max normalization. Multiple machine learning algorithms, including Random Forest, XGBoost, Extra Trees, K-Nearest Neighbors, Voting Classifier, and Stacking Classifier, are implemented for comparative evaluation. Experimental results demonstrate that the Stacking Classifier achieves superior performance with 95.4% accuracy, 95.7% precision, 95.4% recall, 95.5% F1- score, and 99.2% ROC–AUC, indicating strong discriminative capability in identifying anomalous transactions. Model interpretability is enhanced using LIME and SHAP to provide local and global explanations of prediction outcomes. Furthermore, a deployment interface is developed using Flask with SQLite integration to enable secure user authentication and real-time transaction risk prediction. The proposed framework ensures robust, interpretable, and scalable financial anomaly detection suitable for practical applications.