Scholarlink

<p>Technological advancements have enabled the digitalization of nearly every aspect of human daily life, including online shopping. This evolution necessitates the use of credit cards to perform online transactions. Consequently, the volume of daily transactions has significantly increased. However, this rise has been accompanied by a surge in credit card fraud, making the detection of fraudulent transactions essential to minimize financial losses for banks. Over the past two decades, intelligent systems based on Machine Learning (ML) algorithms have been widely explored to combat such losses. Although several studies have explored ML techniques for fraud detection, no consensus has been reached on the most effective model. As a promising alternative, ensemble methods; which combine the predictions of multiple individual ML models to make a final decision, have gained significant attention in various research domains. In this paper, we investigate the application of ensemble methods for credit card fraud detection by evaluating eleven different combination rules. The ensemble members were selected from a pool of 15 distinct classification algorithms. The empirical analysis was conducted using a European credit card dataset, incorporating the Mutual Information feature selection technique and the Synthetic Minority Oversampling Technique algorithm to handle class imbalance. Eight performance metrics were employed to evaluate the predictive capabilities of the proposed ensemble models. The results, statistically validated using the Scott-Knott test, demonstrated that ensemble methods yielded better predictive performance compared to their individual constituent models. Among the combination rules tested, the Multilayer Perceptron-based rule achieved the best performance in our empirical study. Furthermore, the study presents an ablation analysis to evaluate the contribution of the preprocessing components, along with a SHAP-based explainability investigation to interpret the behavior of the stacking ensemble.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evaluating Ensemble Combination Rules for Credit Card Fraud Detection: An Empirical Study

  • 期刊论文

  • 发表日期：

    2026年6月2日

查看全文
  • Chaimae Chekira,
  • Rachid Menaoui,
  • Medarhri Ibtissam,
  • Mohamed Hosni,
  • Zakaria Belhaj

摘要

Technological advancements have enabled the digitalization of nearly every aspect of human daily life, including online shopping. This evolution necessitates the use of credit cards to perform online transactions. Consequently, the volume of daily transactions has significantly increased. However, this rise has been accompanied by a surge in credit card fraud, making the detection of fraudulent transactions essential to minimize financial losses for banks. Over the past two decades, intelligent systems based on Machine Learning (ML) algorithms have been widely explored to combat such losses. Although several studies have explored ML techniques for fraud detection, no consensus has been reached on the most effective model. As a promising alternative, ensemble methods; which combine the predictions of multiple individual ML models to make a final decision, have gained significant attention in various research domains. In this paper, we investigate the application of ensemble methods for credit card fraud detection by evaluating eleven different combination rules. The ensemble members were selected from a pool of 15 distinct classification algorithms. The empirical analysis was conducted using a European credit card dataset, incorporating the Mutual Information feature selection technique and the Synthetic Minority Oversampling Technique algorithm to handle class imbalance. Eight performance metrics were employed to evaluate the predictive capabilities of the proposed ensemble models. The results, statistically validated using the Scott-Knott test, demonstrated that ensemble methods yielded better predictive performance compared to their individual constituent models. Among the combination rules tested, the Multilayer Perceptron-based rule achieved the best performance in our empirical study. Furthermore, the study presents an ablation analysis to evaluate the contribution of the preprocessing components, along with a SHAP-based explainability investigation to interpret the behavior of the stacking ensemble.