A Machine Learning Method with Hybrid Feature Selection for Improved Credit Card Fraud Detection

Abstract

With the rapid developments in electronic commerce and digital payment technologies, credit card transactions have increased significantly. Machine learning (ML) has been vital in analyzing customer data to detect and prevent fraud. However, the presence of redundant and irrelevant features in most real-world credit card data degrades the performance of ML classifiers. This study proposes a hybrid feature-selection technique consisting of filter and wrapper feature-selection steps to ensure that only the most relevant features are used for machine learning. The proposed method uses the information gain (IG) technique to rank the features, and the top-ranked features are fed to a genetic algorithm (GA) wrapper, which uses the extreme learning machine (ELM) as the learning algorithm. Meanwhile, the proposed GA wrapper is optimized for imbalanced classification using the geometric mean (G-mean) as the fitness function instead of the conventional accuracy metric. The proposed approach achieved a sensitivity and specificity of 0.997 and 0.994, respectively, outperforming other baseline techniques and methods in the recent literature.

1. Introduction

Over the years, electronic payments (e-payments) have been the most common payment option due to technological advancements and the development of several electronic funding methods [1]. E-payment systems are essential to the present competitive financial sector and are mostly performed using credit cards [2]. The introduction of credit cards has resulted in convenient and seamless e-payments. A recent study stated that in the second quarter of 2021, Mastercard and Visa issued 1131 million and 1156 million cards, respectively [3]. However, the rise of credit card usage globally has increased the fraud rate, affecting consumers and merchants [4]. For instance, a report stated that financial losses due to credit and debit cards are among the leading causes of losses in the financial sector [3]. Therefore, developing efficient credit card fraud-detection systems is necessary to reduce such losses.

Machine learning algorithms have been widely employed to detect credit card fraud [5,6,7]. Meanwhile, there have been enormous datasets with very high dimensions due to the advent of big data and the Internet of Things (IoT) [8,9]. Furthermore, some features in these datasets might be redundant or less significant to the response variable. Using such features for machine learning could increase the complexity of the model and lead to overfitting [10]. Therefore, to handle the high dimensionality issue, an approach containing dimensionality reduction, such as feature selection, is necessary to obtain valuable insights and make accurate predictions [11].

Feature-selection techniques aim to identify the most important attributes needed to develop a well-performing machine learning model [12,13], ensuring improved classification performance and reduced computational complexity by removing irrelevant and redundant features. Feature selection techniques are usually grouped into three methodological groups: filters, wrappers, and embedded methods [10,14]. The internal workings and configuration of the various feature-selection methods make them suitable for different applications. Filter methods employ attribute ranking to determine the most informative features. Features that attain scores above a given threshold are selected, and those below the threshold are discarded. After identifying the most important features, they can be fed as input to the learning algorithm. Filter methods vary from wrapper and embedded methods as they are not dependent on a classifier and are, therefore, independent of the classifier’s bias [15].

However, wrapper methods use an ML classifier’s performance as the evaluation metric in selecting the most relevant feature set. Wrapper methods usually lead to better classification performance than filter techniques because the feature-selection procedure is optimized for the chosen classification algorithm [16,17]. Generally, wrapper methods employ a search strategy to identify the candidate subsets. The classifier’s performance on the various feature subsets is measured, and the subset that leads to the highest performance is selected as the most informative subset. Examples of wrapper-based feature selection techniques include the Boruta algorithm, forward selection, backward elimination, and the genetic algorithm. Embedded methods select the features that enhance the model’s performance during training. The feature selection is incorporated into the learning procedure [13]. Unlike wrapper methods, this type of feature selection aims to reduce the time used in training different subsets. Embedded methods include random forest, decision tree, gradient boosting, elastic net, and LASSO [10].

Meanwhile, the GA wrapper is an effective method for feature selection, with applications in diverse domains, including natural language processing (NLP) [18], fraud detection [19], sentiment analysis [20], and medical diagnosis [21]. This study proposes a hybrid feature-selection approach, combining the IG-based filter and GA-based wrapper techniques. The main contributions and objectives of the work include the following:

• Using the information gain technique for initial feature selection to rank the features in the credit card dataset, only the top-ranked features are fed into the GA wrapper to reduce the search space and enhance the classification performance.
• Secondly, the GA wrapper is employed to select the best feature subset that results in optimal classification performance, and the ELM is employed as the learning algorithm in the GA wrapper.
• Additionally, this study employs the G-mean as the fitness function in the GA wrapper instead of the conventional accuracy evaluation criterion, ensuring the recognition rate of the minority samples is considered and improved.

The rationale behind this approach is that the initial IG-based feature selection and ELM’s ability to produce promising performance while converging faster than traditional neural networks could reduce the computational complexity of the GA and improve the classification performance. The ELM is chosen as the learning algorithm in the GA wrapper because it converges far more rapidly and achieves higher generalization performance than conventional neural networks. At the same time, its learning process is thousands of times quicker than neural networks trained via backpropagation [22]. Furthermore, for convenience, the proposed hybrid approach is called IG-GAW. It would be compared with the conventional ELM classifier, an ELM classifier with IG-based feature selection (IG-ELM), the GA wrapper (GAW), and well-performing methods in related literature.

The rest of this paper is structured as follows: Section 2 presents related works, focusing on feature-selection methods in the literature. Section 3 discusses the dataset and algorithms used in this study. The proposed credit card fraud-prediction approach is introduced in Section 4. Section 5 presents the results, while Section 6 concludes the paper and provides appropriate future research directions.

2. Related Works

Recently, ML algorithms have been widely applied for credit card fraud detection [23,24,25]. Researchers have used both traditional ML and deep learning (DL) algorithms to predict credit card fraud efficiently. For example, Alarfaj et al. [26] conducted a study using ML and DL techniques for detecting credit card fraud, while Van Belle et al. [27] employed inductive graph representation learning, Esenogho et al. [28] used a neural network ensemble, and Zhang et al. [29] employed an ensemble classifier based on isolation forest and adaptive boosting.

Some problems encountered when dealing with credit card datasets include high dimensionality and imbalance class [30,31], making it difficult for ML classifiers to learn and make accurate predictions. In addition, high dimensional data often make the learning process complex and computationally expensive, resulting in models with poor generalization ability [32]. Therefore, feature selection is essential in such datasets to reduce the computational burden and enhance the model’s generalization ability. For example, Chaquet-Ulldemolins et al. [33] recorded an increase in the classification performance of ML classifiers after introducing feature selection. Generally, feature-selection methods are useful in applications where the number of features affects the classifier’s performance.

The wrapper feature-selection methods have been widely applied in numerous applications [34,35]. They compute the importance of each feature based on its usefulness when training the ML model. The primary components of a wrapper method are the learning classifier and search strategy. The wrapper technique exists as a wrapper around the learning classifier and uses the same classifier to select the most relevant features. Therefore, a robust learning classifier could enhance the wrapper-based feature selection. Furthermore, the search strategy employed in the wrapper could affect the feature selection, and using the right search strategy for a given application is crucial in obtaining good performance.

Evolutionary search techniques such as genetic algorithms can avoid becoming stuck in local optima. Unlike deterministic algorithms, they can identify reduced feature sets that can effectively represent the original feature set [36]. The GA-based wrapper can easily identify feature redundancy and correlations. In addition, selecting a suitable classifier is vital in developing robust GA wrapper models since the wrapper procedure is tied to the selected classifier’s performance. However, there are specific issues to consider when selecting the classifier. Firstly, the classifier should be able to achieve good classification performance and have excellent generalization ability. Secondly, since the classifier would be used to train numerous subsets, it should have good training speed. Thirdly, the number of features in the various subsets might differ. Therefore, using the same model parameters might not be enough to obtain good performance in all the subsets [37]. Hence, it would be preferred to use a classifier that automatically updates the model parameters for every feature subset to achieve good performance.

Other recent methods for credit card fraud detection include a signal processing framework [38], signal processing on graphs [39], and a deep learning ensemble [40]. In addition, in the literature, several learning algorithms (such as decision tree [41], naïve Bayes [42], SVM [43], and random forest [44]) have been used as the classifier in the GA wrapper. However, these classifiers are not able to consider the issues mentioned above. Therefore, a hybrid wrapper approach that considers all the above-mentioned issues is proposed. The proposed approach employs the IG-based filter feature selection to rank the attributes, and only the top-ranked features would be used as input into the GA wrapper. Meanwhile, the GA wrapper employs the ELM as the learning classifier. The ELM can achieve excellent classification performance and generalization ability with an extremely fast learning speed compared to conventional training methods. Furthermore, unlike traditional neural networks based on backpropagation algorithms, the ELM’s training process is entirely automatic and does not require it to be tuned iteratively.

3. Materials and Methods

3.1. Credit Card Dataset

The European cardholders dataset [45] is used in this study. It is publicly available and comprises 284807 transactions made by European cardholders in September 2013. The dataset has been widely used in different credit card fraud-detection studies [1,6,46]. It contains 492 fraudulent transactions, and the rest are legitimate transactions, i.e., only 0.17% of the dataset belongs to the minority class, and 99.83% belongs to the majority class; hence, the dataset is highly skewed and it is challenging for conventional ML algorithms to learn from the dataset.

Due to privacy concerns, the features in the dataset were anonymized as $V1,V2$, …, and $V28$, except for the “Time” and “Amount” features. The “Time” attribute indicates the seconds elapsed between a transaction and the first transaction in the dataset, whereas the “Amount” indicates the value of the transaction. Meanwhile, the “Class” attribute is the response variable, representing legitimate and fraudulent transactions and having values 0 and 1, respectively.

3.2. Information Gain

The information gain technique, or mutual information, is one of the most used filter criteria. The IG criterion is modeled after the concept of entropy in information theory [47]. The entropy measures the impurity or uncertainty in a group of observations, while information gain computes the decrease in entropy before and after adding an attribute. An attribute with a high IG value is usually preferred to those with low IG values. Assuming X and Y are features in a dataset, the information gain of X, given Y is represented mathematically, is:

$$G\left(X\right|Y)=H(X)-H(X\left|Y\right)$$

(1)

In (1), $H\left(X\right)$ denotes the entropy of X and $H\left(X\right|Y)$ is the conditional entropy for X given Y [48]. Meanwhile, $H\left(X\right)$ and $H\left(X\right|Y)$ can be represented as:

$$H\left(X\right)=-\sum _{x\in X}P\left(x\right)lo{g}_2\left(x\right)$$

(2)

$$H\left(X\right|Y)=-\sum _{x\in X}P\left(x\right)\sum _{y\in Y}P\left(x\right|y)lo{g}_2\left(P\left(x\right|y)\right)$$

(3)

Therefore, given two features X and Z, a response variable Y is more correlated to X than Z if $IG\left(X\right|Y)>IG(Z\left|Y\right)$ [49]. Lastly, the information gain technique considers each feature separately, computes the IG value, and outputs its importance to the response variable.

3.3. Genetic Algorithm

The genetic algorithm, inspired by genetics in biological systems, can perform well in high-dimensional feature-selection problems because of its robustness, making it suitable for credit card fraud detection. Given a set of candidate features (called the population), the GA finds the optimal solution via a series of iterative genetic operations. It is superior to most traditional search methods in three main areas: firstly, the GA conducts a parallel search all over the population of solutions; instead of optimizing its parameters, the GA uses chromosomes, an encoded form of a possible solution, to achieve faster convergence. Lastly, the GA employs a fitness value to identify a potential solution. The flowchart of the GA is shown in Figure 1.

The genetic operators include crossover, mutation, and selection [50]. The crossover operator determines how the chromosomes are combined to obtain a new chromosome for the next generation [51]. In the GA, a population of candidate solutions (also known as individuals) to the optimization problem is evolved to obtain the optimal solution. The various candidate solutions have certain attributes called chromosomes that can be mutated; meanwhile, these solutions are represented in 0 and 1 binary strings, indicating whether the corresponding attribute has been selected or not. The mutation operator ensures some bits of the chromosomes are flipped randomly based on probability. The selection step involves choosing chromosomes based on their fitness score for further processing.

A common termination criterion is running the algorithm for a specified number of times [52]. Therefore, the algorithm ends after the specified number of iterations, outputting the optimal solution identified after going through all the generations.

3.4. Extreme Learning Machine

The ELM was developed by Huang et al. [50] to fix the slow learning speed of feedforward neural networks. The authors attributed the slow learning speed to the use of gradient descent-based learning algorithms for training neural networks and how such algorithms iteratively tune the neural network parameters [22]. The ELM has excellent generalization ability with extremely fast learning.

Unlike traditional neural networks, the hidden layer parameters of the ELM are randomly generated without being iterative tuned [37], thereby reducing the learning procedure to just estimating the optimal output weights $\beta$. For a given dataset ${(x_j,t_j)}_{(j=1)}^N$, where N is the number of instances and the SLFN has L hidden nodes, the activations function $g\left(x\right)$ can be mathematically represented as:

$$\sum _{i=1}^L{\beta }_{i}g(w_i.x_j+b_i)=t_j$$

(4)

where $t_j$ is the output of the network, $w_i={[w_{i1},\cdots ,w_{in}]}^T$ represent the input weight, $b_i$ denotes the bias of the $i-th$ hidden node, ${\beta }_i={[{\beta }_{i1},\cdots ,{\beta }_{im}]}^T$ denotes the weight vector linking the $i-th$ hidden node with the output nodes, and $w_i.x_j$ represents the inner product of $w_i$ and $x_j$ [53]. Equation (4) can be rewritten in the compact matrix form as

$$H\beta =T$$

(5)

where H represents the output matrix of the hidden layer and its mathematical formulation, presented by Huang et al. [22], is expressed as

$$\mathrm{H}(w_1,\dots ,w_N,b_1,\dots ,b_N,x_1,\dots ,x_N)={\left[\begin{array}{ccc}g(w_i.x_i+b_i)& \cdots & g(w_N.x_i+b_i)\\ ⋮& \ddots & ⋮\\ g(w_1.x_N+b_1)& \cdots & g(w_N.x_N+b_N)\end{array}\right]}_{N\times N}$$

(6)

$$\beta ={\left[\begin{array}{c}{\beta }_1^T\\ ⋮\\ {\beta }_N^T\end{array}\right]}_{N\times m}and{\left[\begin{array}{c}{t}_1^T\\ ⋮\\ t_N^T\end{array}\right]}_{N\times m}$$

(7)

4. Proposed Credit Card Fraud-Detection Approach

This study employs a hybrid feature-selection method, combining the IG-based filter and GA-based wrapper methods. Firstly, the IG technique ranks the attributes to identify the most significant attributes, and the threshold value is obtained by computing the standard deviation [1] of the IG values. Therefore, features with IG values greater or equal to the threshold are selected, while those below the threshold are discarded. The standard deviation has been widely employed in the literature to achieve excellent feature-selection thresholds [54,55,56]. Secondly, the top-ranked features are fed into the GAW, which uses the ELM as the learning algorithm.

The GA aims to identify the best feature subsets from a series of combinations known as generations [57]. Furthermore, after the ELM classifies the chromosomes, they are assigned a fitness value. Generally, the accuracy metric has been used as the standard fitness function [58]. However, it gives equal importance to samples in both majority and minority classes; hence, it is unsuitable for imbalanced classification problems [59].

Since the majority class samples outnumber the minority class, the fitness function will be biased toward the former. Furthermore, since there are more majority class samples, the accuracy metric will return high accuracy values, which could be misleading. Therefore, this study employs the G-mean metric obtained by the chromosomes as the fitness value to handle the imbalanced class. G-mean is a vital metric for imbalanced classification problems, and it considers the classifier’s performance for the majority and minority classes. The G-mean can be represented mathematically as follows:

$$G-mean=\sqrt{\frac{TP}{TP+FN}\times \frac{TN}{TN+FP}}$$

(8)

where true positive (TP) indicates a fraudulent transaction that is correctly predicted by the model, false positive (FP) indicates a legitimate transaction that the model wrongly predicts as fraud, true negative (TN) indicates a legitimate transaction that is correctly identified, and false negative (FN) indicates a fraudulent transaction that is predicted as legitimate. The proposed approach is outlined in Algorithm 1.

Algorithm 1 Proposed IG-GAW

Compute the information gain of the features in the dataset Rank features according to their importance: $F=(f_1>f_2>f_3,\dots ,f_n)$. GAW input: Top-ranked features from Step 2 and class variable C, population size n, elitism rate e, number of iterations k. Begin Initialize population with n random solutions Compute the fitness values for each random solution. for $i=1,\dots ,k$: Select the best individuals with respect to e Generate new offspring based on the GA operators Examine the fitness values of the new individuals. Discard least-fit population individuals in the population end for GAW output: Optimal solution S

The GAW approach starts with randomly initializing the GA’s population, where every candidate feature subset is encoded as a chromosome. The next step involves training different ELM neural networks based on each chromosome, and the fitness value for each feature subset is computed. Thirdly, a new population is generated using genetic operators. The procedure continues until the stopping criterion is obtained, i.e., the maximum number of generations. This stopping criterion has been used extensively in the literature to obtain excellent GA performance [60,61,62]. Meanwhile, the rationale behind this approach is that the GA wrapper would select the best feature subset that would lead to enhanced prediction performance.

5. Results and Discussion

The proposed method’s classification performance and other baseline classifiers are presented and discussed in the section. The machine learning models were implemented using scikit-learn [63], a widely used library for machine learning in Python. Meanwhile, the proposed method is compared with the following baseline classifiers: AdaBoost [64], logistic regression (LR) [65], random forest (RF) [66], SVM [67], and decision tree [68]. Furthermore, the stratified 10-fold cross-validation method is employed to measure the performance of the prediction models. The stratified k-fold technique ensures that the proportion of fraudulent and legitimate instances in the dataset is preserved in each fold, and it is usually more suitable for imbalance classification problems than the k-fold cross-validation method [69].

The following metrics are used to evaluate the performance of the models: sensitivity, specificity, the receiver operating characteristic (ROC) curve, and the area under the ROC curve (AUC). Sensitivity refers to the model’s ability to predict a fraud transaction as fraud. Usually, a highly sensitive model is preferred in fraud detection, as it implies there are no or few false negative predictions [70,71]. Meanwhile, the specificity of a model is its ability to predict non-fraudulent transactions as legitimate. A model with high specificity means there are hardly any false positive predictions [72]. Sensitivity and specificity can be computed as follows:

$$\mathrm{Sensitivity}=\frac{TP}{TP+FN}$$

(9)

$$\mathrm{Specificity}=\frac{TN}{TN+FP}$$

(10)

The ROC curve is used to visualize the performance of binary classifiers. It plots the true positive rate against the false positive rate at various classification thresholds [73]. In contrast, the AUC summarizes the ROC curve, and its value ranges from 0 to 1. An AUC of 0 implies the model’s predictions are all wrong, and a value of 1 implies all the model predictions are correct. The AUC is a crucial metric in imbalance classification problems, such as fraud detection, as it indicates the classifier’s ability to differentiate between the fraud and non-fraud classes.

Furthermore, in line with existing literature, this study estimates the best GA parameters by conducting several trials using different combinations [74,75,76]. The final GA parameters used in this study are outlined in

Table 1. GA parameters.

Parameter	Value
Population size	50
Number of generations	100
Crossover rate	0.6
Mutation rate	0.01
Fitness function	G-mean
Stopping criteria	Max number of generations
Type of mutation	Uniform mutation
Type of crossover	Single point
Parent selection method	Tournament selection
Tournament size	2

.

5.1. Performance of the ELM Classifier with Filter, Wrapper, and Hybrid Feature Selection Methods

The ELM’s performance without feature selection is compared with instances where the ELM classifier is coupled with the filter, wrapper, and hybrid feature-selection techniques. Firstly, the performance of the ELM without feature selection is recorded. Secondly, the credit card features and their IG values are ranked by the IG-based filter technique.

The standard deviation [54] of the IG values is calculated and used as the threshold value to select the most informative features. From

Table 2. Performance of the ELM classifier, filter, wrapper, and hybrid feature-selection methods.

Classifier	Sensitivity	Specificity	AUC	G-Mean
ELM	0.881	0.904	0.900	0.892
IG-ELM	0.936	0.960	0.940	0.947
GAW	0.949	0.962	0.950	0.955
IG-GAW	0.997	0.994	0.990	0.994

, the standard deviation is 0.145. Hence, information gain values above 0.145 are chosen as the essential features employed for training the machine learning model. The features with IG values below 0.145 are removed. Therefore, the top 21 attributes are chosen by the filter approach as the most important features, while the following attributes are removed: $V8,V19,V24,V23,V26,V13,V25,V15,$ and $V22$.

Thirdly, the GA wrapper coupled with the ELM classifier (GAW) is trained, and its performance is recorded. Lastly, the top 21 features selected by the IG technique are used as input to the GAW. The performance from the four scenarios is tabulated in Table 2, i.e., ELM without feature selection, filter-based IG-ELM, wrapper-based GAW, and the hybrid IG-GAW method. In addition, Figure 2 shows the ROC curves and AUC of the various models. Meanwhile, the complete feature set and the features selected by the various feature-selection methods are tabulated in

Table 3. Feature sets from the European cardholders dataset.

Feature-Selection Method	Features
Complete feature set	$V1,V2,V3,V4,V5,V6,V7,V8,V9,V10,V11$, $V12,V13,V14,V15,V16,V17,V18,V19,$ $V20,V21,V22,V23,V24,V25,V26,$ $V27,V28$, Time, Amount, Class
IG	$V1,V2,V3,V4,V5,V6,V7,V9,V10,V11,$ $V12,V14,V16,V17,V18,V20,$ $V21,V27,V28,$ Time, Amount
GAW	$V1,V2,V5,V6,V7,V9,V10,V11,V12,V16,V17,$ $V18,V20,V21,V27,V28$, Time, Amount
IG-GAW	$V1,V2,V5,V6,V7,V9,V11,V12,V16,V17,V18,$ $V20,V21,V27,V28$

.

Table 2 and Figure 2 show that the proposed hybrid IG-GAW obtained the highest sensitivity, specificity, and AUC values of 0.997, 0.994, and 0.990, respectively, outperforming the ELM, the IG-ELM, and IG-GAW.

5.2. Performance Comparison with Baseline Classifiers and Recent Literature

In this section, the proposed hybrid approach is benchmarked with other ML classifiers and methods in the literature. The baseline classifiers were trained with the complete feature set. The performance of the classifiers is tabulated in

Table 4. Performance comparison with other baseline classifiers.

Classifier	Sensitivity	Specificity	AUC	G-Mean
AdaBoost	0.889	0.918	0.900	0.903
LR	0.752	0.916	0.810	0.829
RF	0.869	0.940	0.890	0.904
SVM	0.585	0.827	0.660	0.695
DT	0.590	0.801	0.690	0.688
Proposed IG-GAW	0.997	0.994	0.990	0.994

and visualized in Figure 3. The models obtained relatively high specificity compared to the sensitivity, which implies the model correctly predicted more non-fraud transactions (majority class) than fraud transactions (minority class).

Meanwhile, in credit card fraud detection, like every imbalance classification task, it is more important to predict the minority class samples correctly. However, the proposed method achieved excellent sensitivity and specificity, indicating its robustness in predicting the minority and majority class samples. This enhanced performance could be attributed to using the G-mean as the fitness function rather than the widely used accuracy criterion, ensuring the model’s detection rate on the minority class is enhanced.

Furthermore,

Table 5. Performance comparison with other well-performing methods in the literature.

Reference	Algorithm	Sensitivity	Specificity	AUC
Zhu et al. [77]	Weighted ELM	0.982	-	0.978
Alkhatib et al. [78]	DNN	0.955	-	0.990
Yotsawat et al. [79]	CS-NNE	-	0.936	0.980
Ileberi et al. [19]	GA-RF	72.56	-	0.950
Kalid et al. [80]	DT-NB	0.872	1.000	-
Mrozek et al. [81]	Random forest-SMOTE	0.829	-	0.910
Carta et al. [82]	Stochastic ensemble	0.915	-	0.876
Xie et al. [83]	XGBoost-SMOTE	0.988	-	0.970
Saheed et al. [84]	GA-SVM	0.963	0.963	-
Verma et al. [85]	PSO-based Ensemble model	0.97	-	-
Padhi et al. [86]	RHSO	0.951	-	-
Ganji et al. [87]	DRN-SSPO	0.912	0.902	-
This paper	Proposed IG-GAW	0.997	0.994	0.990

shows the performance of some state-of-the-art methods in the literature. The methods include a weighted extreme learning machine (weighted ELM) [77], a deep neural network (DNN)-based classifier [78], a cost-sensitive neural network ensemble (CS-NNE) [79], a random forest-based genetic algorithm wrapper method (GA-RF) [19], a method that sequentially combines C4.5 and the naïve Bayes model (DT-NB) [80], a model developed using the random forest algorithm with the SMOTE technique (RF-SMOTE) [81], a stochastic ensemble model [82], an XGBoost-SMOTE model [83], a support vector machine (SVM)-based GA wrapper [84], an ensemble model optimized using the particle swarm optimization (PSO) technique [85], a metaheuristic based feature selection algorithm based on rock hyrax swarm optimization (RHSO) [86], and a deep residual network with shuffled shepherd optimization (DRN-SSPO) [87].

Table 5 shows that the proposed IG-GAW demonstrated excellent results compared to the state-of-the-art methods, indicating the proposed method’s robustness. Furthermore, the proposed method is applied to other credit card datasets to show how the method performs in different scenarios. The datasets include the German credit card dataset [88] and the Taiwan credit card dataset [89]. The German dataset has 20 attributes and contains 1000 samples, of which 700 are classified as good and 300 as bad customers, i.e., 70% of the dataset belongs to the majority class, while 30% belongs to the minority class; hence, the dataset is imbalanced. The features in the German dataset and those selected by the various feature selection techniques are shown in

Table 6. Feature sets from the German dataset.

Feature-Selection Method	Features
Complete feature set	Status of existing checking account, duration in month, credit history, purpose, credit amount, savings account, present employment since, installment rate as a percentage of disposable income, personal status and sex, other debtors, present residence since, property, age, other installment plans, housing, number of existing credits at this bank, job, number of dependents, telephone, foreign worker
IG	Status of existing checking account, duration in month, credit history, purpose, credit amount, savings account, present employment since, installment rate as a percentage of disposable income, personal status and sex, other debtors, property, age, other installment plans, housing, number of dependents, foreign worker
GAW	Status of existing checking account, duration in month, credit history, purpose, credit amount, savings account, present employment since, property, age, other installment plans, housing, number of dependents, foreign worker
IG-GAW	Credit amount, status of existing checking account, duration in months, age, credit history, purpose, property, present employment since, and housing

, while

Table 7. Performance comparison using the German dataset.

Classifier	Sensitivity	Specificity	AUC	G-Mean
AdaBoost	0.785	0.892	0.810	0.837
LR	0.688	0.813	0.700	0.748
RF	0.796	0.904	0.830	0.850
SVM	0.649	0.792	0.650	0.716
DT	0.630	0.787	0.640	0.704
ELM	0.704	0.830	0.710	0.763
IG-ELM	0.796	0.903	0.810	0.847
GAW	0.820	0.925	0.860	0.871
Proposed IG-GAW	0.904	0.946	0.910	0.925

and Figure 4 show the performance of the various models.

Meanwhile, the Taiwan dataset has 25 attributes and 30,000 samples, of which 23,364 are classified as good and 6636 as bad customers, i.e., 77.88% of the dataset belongs to the majority class and 22.12% belongs to the minority class. Therefore, the Taiwan dataset is also imbalanced. The features in the datasets and the selected feature sets are shown in

Table 8. Feature sets from the Taiwan dataset.

Feature Selection Method	Features
Complete feature set	ID, LIMIT_BAL, SEX, EDUCATION, MARRIAGE, AGE, PAY_0, PAY_2, PAY_3, PAY_4, PAY_5, PAY_6, BILL_AMT1, BILL_AMT2, BILL_AMT3, BILL_AMT4, BILL_AMT5, BILL_AMT6, PAY_AMT1, PAY_AMT2, PAY_AMT3, PAY_AMT4, PAY_AMT5, PAY_AMT6
IG	SEX, PAY_0, PAY_2, PAY_3, PAY_4, PAY_5, PAY_6, BILL_AMT1, BILL_AMT2, BILL_AMT3, BILL_AMT4, BILL_AMT5, BILL_AMT6, PAY_AMT1, PAY_AMT2, PAY_AMT3, PAY_AMT4, PAY_AMT5, PAY_AMT6
GAW	PAY_0, PAY_2, PAY_4, PAY_5, BILL_AMT1, BILL_AMT2, BILL_AMT3, BILL_AMT4, BILL_AMT5, BILL_AMT6, PAY_AMT1, PAY_AMT2, PAY_AMT3, PAY_AMT4, PAY_AMT5, PAY_AMT6
IG-GAW	BILL_AMT1, BILL_AMT2, BILL_AMT3, BILL_AMT4, BILL_AMT5, BILL_AMT6, PAY_AMT1, PAY_AMT2, PAY_AMT3, PAY_AMT6, PAY_AMT4, PAY_AMT5, PAY_0, and PAY_2

. Meanwhile,

Table 9. Performance comparison using the Taiwan dataset.

Classifier	Sensitivity	Specificity	AUC	G-Mean
AdaBoost	0.870	0.890	0.870	0.880
LR	0.625	0.837	0.640	0.723
RF	0.829	0.914	0.840	0.870
SVM	0.626	0.819	0.650	0.716
DT	0.574	0.773	0.610	0.666
ELM	0.710	0.885	0.730	0.793
IG-ELM	0.874	0.911	0.890	0.892
GAW	0.899	0.920	0.900	0.909
Proposed IG-GAW	0.945	0.961	0.940	0.952

and Figure 5 show the performance of the various methods.

Using the German dataset, the proposed IG-GAW achieved a sensitivity of 0.904, specificity of 0.945, and AUC of 0.910, as shown in Table 7. Meanwhile, the experimental results for the Taiwan dataset are shown in Table 9, and it shows that the proposed method obtained a sensitivity of 0.945, specificity of 0.961, and AUC of 0.940.

5.3. Discussions

Credit card fraud is a huge burden for financial institutions. The advances in e-commerce and digital payment platforms have made credit card fraud more common. This study aimed to utilize machine learning to detect credit card fraud effectively. Meanwhile, since most credit card datasets contain irrelevant attributes that degrade the performance of machine learning algorithms, this study proposed a robust hybrid feature-selection approach comprising filter (information gain technique) and wrapper (genetic algorithm) feature-selection steps, ensuring only the most significant attributes are used for machine learning. This study uses the well-known European Credit Card dataset. The proposed IG-GAW, which uses the ELM as the learning algorithm in the GA wrapper, obtained scores for sensitivity, specificity, AUC, and G-mean of 0.997, 0.994, 0.990, and 0.994, respectively. Additionally, two popular credit risk datasets (German and Taiwan credit datasets) were used to further validate the proposed method’s improved performance.

The proposed IG-GAW outperformed the selected classifiers when trained with the German and Taiwan datasets. The results also showed that the proposed hybrid approach, IG-GAW, achieved superior performance compared to the filter and wrapper methods, i.e., IG-ELM and GAW. Therefore, integrating filter and wrapper techniques in a hybrid setting is a robust approach to detecting credit card fraud. Lastly, the results also showed that introducing feature selection enhanced the ELM’s performance, indicating the importance of effective feature selection.

6. Conclusions

Detecting fraudulent credit card transactions is challenging, and researchers have developed different methods to handle this problem. This study proposed a hybrid approach to enhance the detection rate. The hybrid approach takes advantage of the strength of different feature-selection and ML methods, including information gain, genetic algorithms, and extreme learning machines. The IG technique was employed for initial feature selection, and the top-ranked features served as input to the GA wrapper. Meanwhile, the ELM was used as the learning algorithm in the GA wrapper. The proposed approach outperformed other baseline classifiers and methods in recent literature. Furthermore, the proposed method was applied to two more credit card datasets to validate its performance, and it achieved excellent performance in both datasets, demonstrating its robustness. Therefore, it can be concluded that the proposed hybrid approach is an effective credit card fraud-detection method. Future research work would employ more combinations of evolutionary algorithms and ML-based feature selection methods to enhance new aspects of credit card fraud detection. In addition, future research would explore the potential of obtaining more recent datasets to train ML models.