Enhancing performance of a LENN based CCFD model with a hybrid TOPSIS-ReliefF based feature selection technique

The rapid acceleration of transaction digitization has paved a comfortable path for fraudsters to personate deviant frauds. The loss from credit card fraud on its own is causing massive destruction in the country's revenue. Impelling effective credit card fraud detection (CCFD) models that can handle the changing nature of fraud is one way to minimize this loss. One of the fundamental challenges in developing such models is to identify the relevant input features those are informative in determining the fraud cases. In this article, a hybrid feature selection technique is suggested in order to lift the performance of a Legender Polynomial Neural Network (LENN) based CCFD model. For identifying the optimal features for the model using the hybrid approach, initially the rank and weights of each feature is calculated using a ReliefF algorithm and then the optimal feature size is decided using a TOPSIS based multi criteria decision making approach. The performance of the CCFD model with and without feature selection approach is accessed over two credit card datasets by considering five classification metric

I. Introduction

Nowadays, with the rapid acceleration of transaction digitization, everyone is accustomed to use credit cards for various types of financial transactions. Along with its benefits, it has also paved the path for fraudsters to develop new scams on a regular basis. The loss from credit card fraud alone is wreaking havoc on the nation's revenue. So this serious issue has attracted several researchers for developing more efficient fraud detection system with the foremost priority of predicting fraud cases with high accuracy and reduced false positive cases by streaming the transaction Data. The relevant literature presents several reviews on application of machines learning techniques for revealing fraudulent activities [1]–[3]. A detailed empirical comparison of machine learning and deep learning algorithms with inclusion of data balancing approach is presented in [4] for detecting fraud in the benchmark European card dataset. It becomes abundantly clear that ANN outperforms LR when two data mining techniques, like ANN and LR for CCFD, are compared in [5]. Three CCFD models suggested in [6] are designed using CHFLANN, ANN and DT, where ANN is found to produce better accuracy and elapsed time compared to the other two models. In [7] the performance of ANN based fraud detection is observed after categorizing the transactions from the normal usage patterns using fuzzy c-means clustering. Another LENN based CCFD model trained using back propagation algorithm is suggested in [8]. In [9] a CPNN based CCFD model is developed whose parameters are optimized using a sine cosine algorithm.