An Improved Fraud Detection System Using the LDX Ensemble Machine Learning Technique

Abstract

Credit cards are widely used for fast and convenient cashless transactions. However, the incidence of fraud is increasing due to usage growth. Detecting fraudulent credit card transactions presents a significant challenge for the financial industry. One of the main obstacles is the imbalance between fraudulent and legitimate transactions, as fraud cases are relatively rare, making it difficult for models to identify them accurately. This research proposes a trustworthy fraud detection system using the LDX ensemble machine learning technique based on Logistics Regression, Decision Tree, and XGBoost models. For that purpose, the credit card fraud dataset from Kaggle was examined. The SMOTE and Weight of Evidence encoding approach was used to preprocess the data to improve feature representations and change categorical variables. After that downsampling methods were used to rectify the class imbalance and ensure a balanced dataset. The following machine learning models were used and assessed: Logistic Regression (LR), Random Forest, Gaussian Naïve Base, Decision Tree (DT), Support Vector Machine (SVM). Hyperparameter tuning was applied to each model to enhance performance. The ensemble LDX model's maximum accuracy is roughly 95%, and the outcome is assessed using metrics like precision, recall, and F1 score. This AI-driven approach demonstrates an effective solution for detecting credit card fraud, contributing to enhanced cybersecurity in economic transactions and minimizing business financial risks.