Evaluating the Performance of Two Hybrid Feature Selection Model of Machine Learning for Credit Card Fraud Detection on Classification and Prediction Methods

The hybrid and non-hybrid feature selection model aimed at predicting and classifying whether a
transaction is fraudulent or non-fraudulent using machine learning approaches. The objective of this
study was to use supervised learning framework to differentiate fraudulent and genuine transactions.
The proposed hybrid model utilizes feature selection methods namely; the Principal Component
Analysis (PCA) and PCA-Backward elimination with multiple linear regression and Reduced Error
Pruning Tree classifier (RepT) using Python and WEKA. Through five experiments carried out in this
study, the proposed approach has proven to be effective for eliminating redundant features in the
dataset that does not have significant impact using PCA and Backward elimination method to optimize
the predictive behaviour of the credit card transactions. Our first findings from the experimental results
revealed that the RepT with PCA-Backward-Elimination prediction accuracy 87.37% is higher than
that of multiple linear regressions with PCA 73.35% and PCA-Backward-Elimination 73.34%. Our
second findings also revealed that the RepT with PCA-Backward-Elimination classification accuracy
of 99.9368% is higher than that of multiple linear regressions with PCA 99.9122% and PCABackward-
Elimination 99.9105%. The performance metrics measures on the classification model of
the logistic regression with PCA-Backward-Elimination indicates that the hybrid model negates it with
the expectation of maximization to minimization (99.9122% to 99.9105%) but returns the same results
on the RepT decision tree classification in both cases of 99.9368%. The proposed hybrid feature
selection method with machine learning algorithms outperforms the non-hybrid feature selection
method with machine learning algorithms for classification and prediction accuracy.