Big Data-Driven Deep Learning Models For Real-Time Fraud Detection

Authors

  • Nebras Jalel Ibrahim University of Diyala Author

DOI:

https://doi.org/10.70882/qg6d3k52

Keywords:

Credit Card Fraud Detection; Deep Learning; Imbalanced Classification; Temporal Validation; Big Data; Real-Time Analytics; Reproducible Machine Learning

Abstract

Fraud detection is a rare-event decision problem: a model can report a high overall accuracy while failing to identify the transactions that matter. We therefore evaluated four classifiers under a chronological protocol using the public Credit Card Fraud Detection dataset released by the Machine Learning Group of Université Libre de Bruxelles and Worldline. The dataset contains 284,807 transactions, including 492 labeled fraud cases (0.1727%). The original order was retained and divided into training (70%), validation (15%), and final test (15%) periods. The comparison includes class-weighted logistic regression, a balanced random forest, cost-weighted HistGradientBoosting, and a three-hidden-layer multilayer perceptron (MLP). Each threshold was chosen on the validation period by maximizing F1 and then frozen before testing. On the final period, the MLP produced the highest F1-score (0.817) and MCC (0.823), with 38 true positives, 3 false positives, and 14 false negatives. Random Forest produced the best average precision (PR-AUC = 0.765) and the highest recall (0.750). Thus, the neural model delivered the strongest selected operating point, while the forest ranked transactions slightly better under PR-AUC. The evidence is useful as a controlled benchmark, not as a claim of production readiness.

Downloads

Download data is not yet available.

Downloads

Published

2026-10-16

How to Cite

Big Data-Driven Deep Learning Models For Real-Time Fraud Detection. (2026). Journal of Pure and Applied Sciences (Science Forum), 26(4). https://doi.org/10.70882/qg6d3k52

Similar Articles

281-286 of 286

You may also start an advanced similarity search for this article.