Comparative Analysis of Machine Learning Models for Credit Card Fraud Detection Using SMOTE for Class Imbalance

The exponential growth of e-commerce and the digitization of financial transactions have revolutionized the way people interact with markets and services, bringing unprecedented convenience to everyday life [1]. However, this increased ease, and accessibility has also led to an increase in fraudulent transactions [2], due both to how these systems are used and the tendency of cybercriminals to obtain sensitive information [3]. This type of fraud, recognized as a sophisticated form of financial misconduct, entails the illegal exploitation of credit card information to carry out illicit transactions. Credit card fraud can occur in a variety of ways. It typically happens when someone unlawfully takes a physical card or gains access to its details and uses them without the owner's consent [4]. The methods used to commit fraud are varied and increasingly sophisticated. The most common methods include hacking billing systems, infiltrating online merchants, taking cards without permission, and placing malicious hardware in payment terminals [5].

In addition, criminals use fake identities and various advanced technologies to trick users into giving up their credit card information [6]. One of the most common and effective tactics is phishing, a technique that uses fraudulent emails, SMS messages, or websites designed to deceive cardholders into revealing their personal information [7]. These methods not only put people's financial assets at risk but also pose a significant threat to the security and trust of digital payment systems. As a result, credit card fraud is a multifaceted problem that affects banks and other organizations that offer credit cards [8], as well as consumers and merchants. The complexity of the problem lies in the variety of methods used by criminals, ranging from the physical theft of cards to the creation of counterfeit cards and the online theft of personal information. In addition, the lack of effective security in payment systems and the increasing sophistication of fraudsters make it increasingly difficult to effectively detect and prevent fraud. Consequently, detecting fraudulent credit card transactions represents one of the main challenges faced by financial institutions [9].

This study tackles the problem of fraudulent credit card activity through the use of ML approaches, aiming to determine which model offers the highest accuracy and efficiency in detecting fraud. Seven algorithms were chosen due to their varied theoretical backgrounds, frequent appearance in fraud detection research, and their capability to manage unbalanced datasets. Naive Bayes (NB) represents a probabilistic baseline model that is simple and computationally efficient. Logistic Regression (LR) provides a strong linear classifier often used as a benchmark. k-Nearest Neighbors (KNN) is included for its instance-based learning approach. Decision Trees (DT) offer interpretability and non-linear decision boundaries. Ensemble methods (Random Forest (RF), XGBoost, and AdaBoost) were selected due to their superior performance in complex classification tasks, especially in fraud detection scenarios. By evaluating models from different paradigms (probabilistic, linear, instance-based, tree-based, and ensemble), this comprehensive approach allows for a fair and insightful comparison to determine the most effective model for early and accurate fraud detection.

This study aims to compare several ML algorithms designed to identify fraudulent credit card transactions. It also addresses the frequent issue of class imbalance present in fraud datasets by applying the SMOTE technique. The goal is to identify which models yield the best results according to evaluation metrics such as accuracy, F1 score, and the area under both the ROC and the precision-recall curves.

3.1 Model evaluation process

Figure 1 illustrates the model evaluation process in detail, from the initial loading of the dataset to the final stage of model performance evaluation. This process involves several critical phases. First, the data is loaded and examined, then it goes through a preprocessing phase that involves cleaning the dataset, addressing any missing values, and normalizing the variables when needed. The data is then divided into two separate subsets: one used to train the model and another to evaluate its performance. In this phase, model selection and training are performed using the SMOTE oversampling technique to address class imbalance.

1.png

Figure 1. Diagram illustrating the evaluation process of the selected models

After training, the models are assessed using multiple evaluation metrics to thoroughly understand how well they perform. This workflow ensures a thorough and rigorous evaluation of each model. In addition, this visual representation provides a comprehensive and structured view of the workflow used to evaluate ML models for credit card fraud detection, highlighting each key step in the process and its contribution to the ultimate goal of identifying the efficient model.

3.2 Oversampling selection

SMOTE was selected over other oversampling methods, such as Adaptive Synthetic Sampling (ADASYN), due to its ability to generate synthetic samples in a more controlled and uniform manner. Unlike ADASYN, which may introduce noise by focusing on harder-to-learn instances, SMOTE helps maintain a balanced class distribution without distorting the decision boundary. Its proven reliability in fraud detection tasks makes it a suitable choice for this study.

3.3 Dataset and preprocessing

3.3.1 Data description

For the purposes of this study, the dataset was acquired from Kaggle, a well-known repository for machine learning datasets. It provides comprehensive records of credit card transactions conducted by European cardholders during a specific period in September 2013. The dataset includes transactions from just two days, totaling 284,807 entries, with only 492 flagged as fraudulent [16]. This highlights a significant class imbalance, as fraudulent cases represent just 0.17% of the data. Despite this, the dataset provides a solid foundation for analyzing and developing effective credit card fraud detection models (see Figure 2).

2.png

Figure 2. Number of fraudulent and normal transactions

Figure 3 shows the correlation between the transaction amount and the time at which the transaction took place. Each point represents an individual transaction, some of which are fraudulent and some of which are legitimate, as indicated by the class variable in the dataset. As can be seen, fraudulent transactions tend to cluster in certain ranges of amounts and times, which may indicate suspicious activity. On the other hand, non-fraudulent transactions are more spread out over time and over a wide range of amounts, reflecting the varied and frequent nature of legitimate transactions compared to fraudulent ones.

3.png

Figure 3. Transaction scatter plot

3.3.2 Features

The dataset contains 31 columns, including the variables Time, V1 to V28, Amount, and Class. The input variables are numerical and have been modified through the application of Principal Component Analysis (PCA) to ensure the original data remains confidential. Features V1 to V28 represent the principal components resulting from PCA, while time and amount are the only variables that remained in their original form without being altered by the PCA process. Furthermore, as shown in Table 1, there is no missing data in any of the variables. In addition, a detailed description of what each variable represents is provided.

Table 1. Data characteristics

3.3.3 Data for training and testing

The dataset was divided into two separate subsets, one used to train the model and the other reserved for evaluating its performance. The dataset was divided so that 80% of the data was served for training purposes, while the other 20% was reserved for testing to assess how well the model performs. This splitting strategy ensures that the model can learn effectively during the training phase, and then its predictive ability can be verified using independent data during the testing phase.

3.3.4 Synthetic minority over-sampling technique

Since fraudulent transactions make up only 0.17% of the data, whereas non-fraudulent one’s account for 99.83%, the dataset presents a highly unbalanced distribution between the two classes. This unevenness may lead the model to become biased, often predicting that most transactions are legitimate. To address this issue, the SMOTE technique was used to equalize the class proportions within the dataset. This approach involves producing artificial data points for the underrepresented class (Specifically, the focus is on transactions identified as fraudulent) by interpolating the attributes of existing samples to enhance class balance. By balancing the amount of data between classes, SMOTE helps prevent the model from showing a bias towards a specific class and improves its ability to more accurately identify fraud.

3.4 Machine learning model

A total of seven supervised machine learning models were applied to thoroughly assess the effectiveness of detecting fraudulent credit card activity.

3.4.1 Naive bayes

NB is a probabilistic classification method that relies on statistical theory to determine the most likely class of an object based on its features, under the assumption that they do not influence one another, which is seldom the case in real-world scenarios, hence the term "naïve." It estimates the likelihood of an outcome occurring by taking into account a prior event that has already taken place [17].

3.4.2 Logistic regression

LR is a classification model mainly used for data sets with binary classifications [18], i.e., when the possible outputs are two different categories. Furthermore, the simplicity and effectiveness of logistic regression make it a valuable tool for binary classification problems.

3.4.3 K-nearest neighbors

KNN is a nonparametric classification method used for both classification and regression problems [19]. It functions by locating the k closest data samples relative to a specific input instance and determining the outcome based on the predominant class among those neighbors for classification or the mean of their values for regression.

3.4.4 Decision tree

The DT algorithm is a technique that organizes data in a hierarchical tree structure [20]. It starts from a root node and branches into different paths depending on the characteristics of the dataset [21]. Each node symbolizes a characteristic or trait, while each branch signifies a decision made based on that specific attribute. These decisions lead to classification or prediction results in the leaves of the tree. Interpreting a decision tree is intuitive and allows us to understand how decisions are made based on the characteristics of the data.

3.4.5 Random forest

RF is a powerful ML technique primarily used for classification, although it can also be applied to regression. This approach constructs multiple decision trees using various randomly selected subsets of the training dataset and aggregates their outputs to enhance the model's precision and stability. It is capable of processing both linear and nonlinear data effectively. RF is particularly effective on unbalanced data sets, where classes are not equally represented [22]. By integrating the results of multiple trees, this algorithm reduces overfitting and provides more accurate and reliable predictions.

3.4.6 XGBoost

The XGBoost algorithm is an advanced ML technique based on the boosting method. In simple terms, boosting is the process of combining several weak models to create a strong model. XGBoost specializes in the sequential generation of decision trees. Each new tree is built to correct the errors made by the previous tree. This is done by updating the residual error, which is the error that remains after the previous tree has made its predictions. Thus, each successive tree focuses on improving the predictions by correcting the accumulated residual errors [22]. This iterative and sequential approach allows XGBoost to achieve high accuracy and robust performance in classification and regression tasks.

3.4.7 AdaBoost

The AdaBoost algorithm is a technique designed to improve the accuracy of classification models by combining several weak classifiers into one stronger classifier. Typically, AdaBoost uses decision trees as its base classifiers. The process starts by fitting a weak classifier to the training data. It then chooses the classifier that has the smallest weighted classification error and adjusts the importance of each data point by assigning higher weights to those that were incorrectly predicted. This adjustment is made using a normalization factor to ensure that the sum of all weights is equal to 1. The algorithm repeats this process iteratively, attempting to reduce the classification error at each step until the training data is correctly classified or no further improvement is possible [23]. Through this weighted combination of weak classifiers, AdaBoost builds a strong and efficient model for the classification task.

3.5 Model evaluation

Several performance metrics were used to evaluate the models, providing a complete and detailed view of their performance.

3.5.1 Accuracy

Accuracy is an indicator that measures how well a model makes its predictions. It is a way to evaluate the overall effectiveness of the model in correctly identifying the positive and negative classes in a data set. It is derived by computing the proportion of instances that the model accurately predicts out of the total observations present in the dataset, serving as an indicator of overall model performance [24]. It can be calculated using Eq. (1).

Accuracy $=\frac{T P+T N}{T P+F N+T N+F P}$                        (1)

where, TP stands for the count of positive cases that the model successfully recognized. Conversely, FP indicates the number of negative cases that the model wrongly marked as positive. Likewise, FN represents positive cases that were inaccurately identified as negative, and TN corresponds to the quantity of negative cases that the model correctly classified [25].

3.5.2 Recall

Recall is a measure of how well a classifier can correctly identify instances of the positive class. In other words, it evaluates how often the classifier correctly predicts a positive label when the data is actually positive [25]. It can be calculated using Eq. (2).

Recall $=\frac{T P}{T P+F N}$                         (2)

3.5.3 F1 score

The F1 score is an indicator of the precision of a model, calculated as the harmonic mean between accuracy and sensitivity. Its value can be as high as 1, indicating ideal accuracy and sensitivity, or as low as 0. In other words, the F1 score evaluates both the precision and the robustness of a model [26]. It can be calculated using Eq. (3).

$F 1$ Score $=\frac{2 *( { precision } *\,{ recall })}{( { precision }+ \, { recall })}$                        (3)

3.5.4 Receiver Operating Characteristic

The Receiver Operating Characteristic (ROC) curve is a visual tool that assesses how well a classification model performs by examining different threshold values. It shows the trade-off between the true positive rate, indicating how many actual positive instances were correctly detected, and the false positive rate, which denotes the percentage of negative instances that were mistakenly labeled as positive. Lowering the classification threshold increases the sensitivity of the model, resulting in a higher proportion of cases being classified as positive, including both true positives and false positives [27].

3.5.5 Area under the curve

The area under the curve (AUC) is a key metric derived from the ROC curve that helps assess how well a classification model can differentiate between classes. When the AUC value is high, it means the model performs well in separating positive instances from negative ones with greater accuracy. Its value ranges between 0 and 1, where 0 indicates a model that classifies everything incorrectly and 1 indicates a perfect model that classifies everything correctly [28]. As the AUC value approaches 1, the model demonstrates a stronger capability to accurately distinguish between different classes.

3.5.6 Area under the precision recovery curve

The area under the precision-recall curve (AUPRC) is a key measure used to evaluate how well classification models perform, especially when dealing with datasets where classes are unevenly distributed. This metric reflects how precision and recall trade off against each other as the decision threshold changes. With values ranging from 0 to 1, the AUPRC quantifies the space beneath the curve that plots precision versus recall. A higher AUPRC score signifies a more effective model, capable of accurately distinguishing between positive and negative cases. This measure is vital for fine-tuning models to achieve the best possible balance between precision and recall, ultimately enhancing overall classification accuracy [12].

3.5.7 Confidence interval estimation

To assess the stability of the performance metrics, 95% confidence intervals were calculated using bootstrap resampling. For each model, 1,000 resamples of the test set predictions were generated with replacement, and confidence intervals were derived from empirical distributions. This approach ensures a reliable estimate of model performance variability.

This study evaluated the effectiveness of multiple supervised machine learning algorithms in detecting fraudulent credit card transactions. While previous research has investigated the application of machine learning techniques for recognizing fraudulent credit card transactions, it has generally concentrated on assessing individual models or a limited set of them, without offering a broad comparison using consistent evaluation criteria across multiple algorithms. The lack of head-to-head comparisons makes it difficult to determine the most effective model for this specific task.

The study findings revealed that, out of all the models tested, Random Forest and XGBoost demonstrated notably superior performance. The RF model achieved a remarkably high accuracy of 0.999, an F1 score of 0.872, an AUC of 0.978 and an AUPRC of 0.871, while XGBoost also showed outstanding performance with an accuracy of 0.999, an F1 score of 0.837, an AUC of 0.983 and an AUPRC of 0.867. These AUPRC values, reflecting a strong harmony between precision and recall, highlight how effective both models are at identifying fraudulent credit card activity. The outcomes indicate that Random Forest and XGBoost are top contenders for this application, thanks to their strength in reducing incorrect classifications of both fraudulent and legitimate transactions.

When comparing the outcomes of this research with those reported in earlier works that utilized comparable datasets for identifying fraudulent credit card activity, the Random Forest and XGBoost models emerge as particularly strong performers. They excel in key metrics such as accuracy (99.9%), F1 score (0.872 and 0.837, respectively), and AUPRC (0.871 and 0.867, respectively). These models significantly outperform the models evaluated by Arora et al. [29] and Vengatesan et al. [30], who reported lower accuracies and F1 scores for KNN and NB. Compared to the ensemble-based soft-voting approach of Azim Mim et al. [10], which achieved an accuracy of 0.9996 and an F1 score of 0.8764, the models evaluated in the present study, such as RF and XGBoost, show a competitive performance, although slightly inferior in AUPRC. Furthermore, the results of RF and XGBoost also outperform those of the ANN evaluated by Asha and Suresh Kumar [31] and the LR and Voting Classifier models of Alenzi and Aljehane [14].

These findings underscore the effectiveness of combining oversampling techniques like SMOTE with robust evaluation metrics such as AUC-ROC and AUPRC. They confirm the strong performance of RF and XGBoost in handling class imbalance and detecting fraudulent transactions. However, despite their high predictive performance, both models have limitations. One notable drawback of RF and XGBoost is their computational complexity, especially during training with large datasets or extensive hyperparameter tuning. These models can require significant processing time and memory resources, which may hinder real-time deployment in high-frequency transaction systems. Furthermore, they are often regarded as "black-box" models, making them less interpretable than simpler algorithms like LR or DT. This reduced interpretability can pose challenges in financial environments where regulatory compliance and explain ability of automated decisions are critical.

Another limitation of this study lies in the age of the dataset used, which dates back to 2013. Given the evolving nature of fraudulent behavior and transaction patterns, it is likely that newer fraud strategies have emerged in the past decade. Therefore, future work should aim to validate model performance using more recent datasets to ensure the continued relevance and applicability of the findings.