Year: 2025
Type: article
Abstract: As fraudulent activities become more advanced and rule-based systems reach their limitations, detecting financial fraud puts a big strain on financial institutions. In this thesis, we use Sparkov data and AI to see how effective it is at spotting fraudulent credit card activity, given that 0.57% of the transactions are fraudulent. The study tests Logistic Regression, Random Forest, XGBoost, Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) against a rule-based baseline, checking their accuracy, precision, recall, F1-score and ROC-AUC. The process goes from preprocessing with SMOTE, developing the features, choosing proper hyperparameters and concluding with thorough evaluation. They show that XGBoost and Logistic Regression achieve the best results, reaching recalls of 0.8993 and 0.8844 and ROC-AUC scores of 0.9656 and 0.9624, which beat the baseline of recall (0.6000) and ROC-AUC (0.7500). Yet, the low precision observed suggests there are a lot of false positives, which impacts how the algorithm is used. Although CNN and LSTM score well with precision (0.9699 and 1.0) and recall (0.0751 and 0.0061), since recall is low, both methods cannot be used. It was found that city, amount, job, category and gender have a high impact, which could lead to bias problems in the system. The study advises that XGBoost should be applied for fraud detection and outlines how to improve the model’s results. What they do is increase the effectiveness of fraud detection and give financial institutions a solid structure to address economic, ethical and regulatory factors.
Author: Shahbaj Ahmad
Institution: IU International University of Applied Sciences
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4413416023