Scholarlink

<p>The increasing number of automobile insurance claims has made fraudulent activities a significant challenge for the insurance sector. Effective detection of automobile insurance fraud is essential for minimizing financial losses and preserving the credibility of insurance providers. This study proposes a stacking ensemble-based machine learning (ML) framework integrated with explainable artificial intelligence (XAI) to improve predictive performance and interpretability. A dataset comprising 40 attributes sourced from Mendeley Data is used, which exhibits an inherent class imbalance characteristic commonly observed in fraud detection problems. After preprocessing, feature selection is performed using random forest (RF) feature importance, resulting in the top 20 most relevant features. Seven ML techniques, including support vector machine (SVM), <i>K</i>-nearest neighbors, logistic regression (LR), decision tree (DT), RF, gradient boosting, and AdaBoost, are evaluated on the reduced feature set, with LR and SVM achieving the highest individual accuracy of 0.82. To further enhance predictive performance, stacking ensemble models are developed, where the DT–SVM-based stacking configuration achieves the best performance with an accuracy of 0.84 and an AUC of 0.80. The performance of the proposed framework is evaluated using both cross-validation and a separate test set to ensure robustness and generalization capability. Additionally, SHAP analysis is employed to interpret model predictions, identifying key contributing factors such as <i>property claim</i>, <i>incident severity</i> (minor damage and total loss), <i>vehicle claim</i>, and <i>injury claim</i>. These findings demonstrate that the proposed approach improves fraud detection performance while providing interpretable insights to support transparent decision-making.</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

Explainable Ensemble Learning Framework for Accurate Detection of Automobile Insurance Fraud

  • 期刊论文

  • 发表日期：

    2026年5月30日

查看全文
  • Md Amir Hamja,
  • Md. Ziaul Hassan

摘要

The increasing number of automobile insurance claims has made fraudulent activities a significant challenge for the insurance sector. Effective detection of automobile insurance fraud is essential for minimizing financial losses and preserving the credibility of insurance providers. This study proposes a stacking ensemble-based machine learning (ML) framework integrated with explainable artificial intelligence (XAI) to improve predictive performance and interpretability. A dataset comprising 40 attributes sourced from Mendeley Data is used, which exhibits an inherent class imbalance characteristic commonly observed in fraud detection problems. After preprocessing, feature selection is performed using random forest (RF) feature importance, resulting in the top 20 most relevant features. Seven ML techniques, including support vector machine (SVM), K-nearest neighbors, logistic regression (LR), decision tree (DT), RF, gradient boosting, and AdaBoost, are evaluated on the reduced feature set, with LR and SVM achieving the highest individual accuracy of 0.82. To further enhance predictive performance, stacking ensemble models are developed, where the DT–SVM-based stacking configuration achieves the best performance with an accuracy of 0.84 and an AUC of 0.80. The performance of the proposed framework is evaluated using both cross-validation and a separate test set to ensure robustness and generalization capability. Additionally, SHAP analysis is employed to interpret model predictions, identifying key contributing factors such as property claim, incident severity (minor damage and total loss), vehicle claim, and injury claim. These findings demonstrate that the proposed approach improves fraud detection performance while providing interpretable insights to support transparent decision-making.