Transaction fraud detection using machine learning

All This paper has described the complete design, development, and evaluation of a machine learning-based Transaction Fraud Detection System, built around a Random Forest ensemble classifier trained on the PaySim synthetic financial transaction dataset. The work set out to address three interrelated challenges that consistently complicate fraud detection in practice: the severe imbalance between fraudulent and legitimate transactions, the need for real-time prediction performance, and the requirement for interpretable risk assessments that non-specialist users can act upon. Each of these challenges has been addressed through specific, deliberate design choices. Class imbalance was corrected using SMOTE applied exclusively to the training partition, preventing synthetic samples from contaminating test set evaluations. Feature engineering introduced two balance-error derived attributes — errorBalanceOrig and errorBalanceDest  that encode domain knowledge about expected account behaviour and proved to be among the most discriminative predictors in the final model. Real-time performance was achieved through a Flask-based REST API serving a serialised model with minimal per-request overhead, delivering predictions in an average of 145 milliseconds. Interpretability was addressed through an interactive web frontend that presents fraud probability scores, confidence indicators, and visual risk summaries in a form accessible to financial analysts without machine learning backgrounds. The trained model achieves an overall accuracy of 82.2% with approximately 76% recall and 78% precision for the fraud class. These figures represent a substantial improvement over the near-zero fraud recall that naive majority-class classifiers produce on this dataset, and they confirm that supervised machine learning with appropriate preprocessing can deliver practically useful fraud detection capabilities. The full testing regimen — unit, integration, and system-level  achieved a 100% pass rate across 24 distinct test cases, validating the correctness and robustness of the implementation. Taken together, the results support a broader conclusion: machine learning-based fraud detection is not merely a theoretical proposition but a workable, deployable solution to a problem that continues to cause significant financial harm. The adaptability that rule-based systems lack  the ability to learn new fraud patterns from data rather than relying on manually updated rules  is precisely what the digital financial environment demands. The present system demonstrates that this adaptability can be delivered at scale, in real time, and in a form that integrates naturally into existing web-based financial infrastructure.

[1]Lopez-Rojas, E. A., Elmir, A., and Axelsson, S. (2016). PaySim: A Financial Mobile Money Simulator for Fraud Detection. The 28th European Modeling and Simulation Symposium, EMSS, Larnaca, Cyprus. [2] Breiman, L. (2001). Random Forests. Machine Learning, 45(1), 5–32. Springer, Netherlands. DOI: 10.1023/A:1010933404324 [3] Chawla, N. V., Bowyer, K. W., Hall, L. O., and Kegelmeyer, W. P. (2002). SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial Intelligence Research, 16, 321–357. [4] Pedregosa, F. et al. (2011). Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research, 12, 2825–2830. [5] Géron, A. (2022). Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow. 3rd Edition. O'Reilly Media, Sebastopol, CA. [6] Flask Documentation. (2023). Flask 2.x — Pallets Projects. https://flask.palletsprojects.com [7] Association of Certified Fraud Examiners. (2022). Report to the Nations: 2022 Global Study on Occupational Fraud and Abuse. ACFE, Austin, TX. [8] Reserve Bank of India. (2023). Annual Report on Payment and Settlement Systems. RBI Publications, Mumbai. [9] Mitchell, T. (1997). Machine Learning. McGraw-Hill, New York. [10] Gama, J., and Brazdil, P. (2000). Cascade Generalization. Machine Learning, 41(3), 315–343.