Hybrid Unsupervised Machine Learning for Insurance Fraud Detection - PCA-XGBoost-LOF and Isolation Forest

Aug 5, 2026

Insurance fraud threatens insurers’ financial stability by creating substantial economic losses, motivating more effective detection approaches. The study introduces a hybrid unsupervised machine learning framework that combines Principal Component Analysis (PCA), Extreme Gradient Boosting (XGBoost), Local Outlier Factor (LOF), and Isolation Forest to leverage complementary strengths across models. Experiments on a real-world insurance dataset report 92% detection accuracy with 92% precision and 96% recall. Results indicate the proposed approach outperforms state-of-the-art methods by increasing detection rates while reducing false positives. The work supports practical development of fraud detection systems that help insurers minimize losses and improve profitability.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages19

Reading Time48 min

AccessFree to download

Introduction

Insurance fraud and detection limitations

Hybrid unsupervised learning approaches

Motivation for PCA, XGBoost, LOF, and Isolation Forest

Journal of Information Systems and Informatics Vol. 7, No. 1, March 2025 e-ISSN: 2656-4882 p-ISSN: 2656-5935 DOI: 10.51519/journalisi.v7i1.958 Published By DRPM-UBD Hybrid Unsupervised Machine Learning for Insurance Fraud Detection: PCA-XGBoost-LOF and Isolation Forest Natsai Chapwanya1, Karikoga Norman Gorejena2 1Machine Learning Research Focus Area, Faculty of Economic and Management Sciences, North West University, Mafikeng Campus Private Bag X2046, Mmabatho 2735, South Africa 2 Faculty of Economic and Management Sciences, North West University, Mafikeng Campus Private Bag X2046, Mmabatho 2735, South Africa [Email :](Email :1 nchapwanya@gzu.ac.zw)[1](Email :1 nchapwanya@gzu.ac.zw)[ nchapwanya@gzu.ac.zw](Email :1 nchapwanya@gzu.ac.zw), [2](2 Koga.gorejena@nwu.ac.za)[ Koga.gorejena@nwu.ac.za](2 Koga.gorejena@nwu.ac.za) Abstract Insurance fraud poses a significant threat to the financial stability of insurance companies, resulting in substantial economic losses. To combat this issue, this study proposes a novel unsupervised machine learning hybrid algorithm, integrating Principal Component Analysis (PCA), Extreme Gradient Boosting (XGBoost), Local Outlier Factor (LOF), and Isolation Forest. This hybrid approach aims to improve the detection accuracy of insurance fraud by combining the strengths of each individual algorithm. Experimental results a realworld insurance dataset demonstrate a detection accuracy of 92%, precision of 92% and recall of 96% . Our experimental results demonstrate that the proposed hybrid algorithm outperforms existing state-of-the-art methods, achieving a higher detection rate and reducing false positives. This research contributes to the development of effective insurance fraud detection systems, ultimately helping insurance companies to minimize financial losses and improve their overall profitability. Keywords: Insurance Fraud Detection; Hybrid Machine Learning; Unsupervised Learning; Anomaly Detection; Principal Component Analysis (PCA) 1. INTRODUCTION Insurance fraud can be classified into different types, including auto insurance fraud which is the focus of this study. Machine learning techniques have been used to detect fraud, but most studies use supervised learning methods that require labelled data. Hybridization of unsupervised learning algorithms has emerged as a promising approach for fraud detection, offering several advantages over traditional rule-based systems and individual unsupervised learning algorithms. Previous and current literature suggests that hybrid unsupervised machine learning approaches can significantly improve the accuracy, efficiency, and scalability of fraud detection systems [1] . Hybridization has the potential to significantly improve the accuracy and efficiency of anomaly detection [2] . Unsupervised learning algorithms are a type of machine learning algorithm that can learn from data without being given any labels [3] . 941 This work is licensed under a Creative Commons Attribution 4.0 International License. p-ISSN: 2656-5935 [http://journal-isi.org/index.php/isi](http://journal-isi.org/index.php/isi) e-ISSN: 2656-4882 Hybrid machine learning algorithms combine two or more different machine learning algorithms to improve performance. A number of studies have shown that hybrid machine learning algorithms can improve the accuracy and efficiency of anomaly detection in unsupervised learning scenarios. For example, a study by [4] found that a hybrid ensemble method outperformed individual unsupervised machine learning algorithms for anomaly detection in network traffic data. Another study by [5], found that a hybrid stacked model outperformed individual unsupervised machine learning algorithms for anomaly detection in financial data. Traditional methods for detecting insurance fraud typically involve manual review of claims and spot checks by insurance investigators [6]. In a manual review, claims are reviewed by insurance investiga

FAQs

Why does the study focus on unsupervised machine learning for insurance fraud detection?

Insurance fraud detection often relies on labeled data in supervised methods, which may be unavailable. Unsupervised learning can learn patterns without labels, and hybridization can improve detection effectiveness.

What hybrid models are combined in the proposed method?

The framework integrates Principal Component Analysis (PCA), Extreme Gradient Boosting (XGBoost), Local Outlier Factor (LOF), and Isolation Forest to improve anomaly detection performance.

How effective is the proposed hybrid algorithm based on experimental results?

Experiments on a real-world insurance dataset report 92% accuracy, 92% precision, and 96% recall, with improved detection rate and reduced false positives compared with existing methods.

Aug 5, 2026

Upload by River Wang