A COMPARATIVE AND HYBRID ENSEMBLE FRAMEWORK FOR CREDIT CARD FRAUD DETECTION USING MACHINE LEARNING AND DEEP LEARNING TECHNIQUES

Abstract

Detecting credit card fraud is a critical issue within finance due to the problems associated with the extreme level of class imbalance, the continuously changing be haviour of fraudsters, and the need to make real-time decisions. This thesis aims to provide a contribution to improving machine learning’s application for detecting credit card fraud. The first of these papers provides a structured comparative analysis of su pervised, unsupervised, deep learning, ensemble and hybrid algorithm-based methods for detecting credit card fraud. For the purpose of comparison, algorithms such as logistic regression, random forest, support vector machine, gradient boosting, autoen coder, and isolation forest were compared according to different criteria such as the characteristics of datasets employed, approaches to class imbalance through SMOTE and ADASYN, and multiple performance metrics (i.e. Precision, Recall, F1-Score, and AUC). The second contribution presents a new threshold-tuned hybrid ensemble methodology based on Random Forest, LightGBM and Autoencoder utilizing no syn thetic oversampling technique. This study used an exhaustive search technique for optimising the weights assigned to each classifier and the decision threshold in order to maximise the F1-Score on the original data with the class imbalances. The frame work was tested using the datasets for credit card fraud detection available through the Kaggle. The results of this method were impressive, yielding overall performance metrics: 99.96% accuracy, 0.9318 precision, 0.8367 recall, 0.8817 F1 score and 0.9756 ROC-AUC with this method of approach outperforming all individual classifiers.