BỘ GIÁO DỤC VÀ ĐÀO TẠO ĐẠI HỌC KINH TÉ THÀNH PHÓ HÒ CHÍ MINH BÁO CÁO TÔNG KÉT ĐÈ TÀI NGHIÊN cứu KHOA HỌC THAM GIA XÉT GIÁI THƯỞNG “NHÀ NGHIÊN cứù TRẺ UEH” NÀM 2024 ANALYZING FRAUD RATE BASED ON INSURANCE CONTRACT REPORTS USING META-DEEP STACKING APPROACH Thuộc nhỏm chuyên ngành : 4. 1 ABSTRACT Insurance fraud, particularly in automobiles, presents substantial challenges for insurance companies worldwide. Fraudulent activities by policyholders, such as document falsification and evidence fabrication, aim to deceive insurers and unlawfully obtain funds, resulting in significant financial losses. However, many insurers rely solely on traditional financial assessment methods, which may be insufficient in detecting fraud, especially in cases involving sophisticated schemes or high claim volumes.
This study seeks to develop an effective fraud detection model tailored to the characteristics of real-world data using machine learning algorithms. The dataset comprises 1000 insurance claims related to car collisions across seven US states in 2015. Results indicate that employing an Early Fusion deep learning model, integrating Decision Tree and Random Forest, mitigates the limitations of traditional models. The findings demonstrate that the proposed model outperforms previous traditional models and enhances fraud detection capabilities.
Building on these findings, the objective is to improve fraud detection within the Vietnamese insurance industry, thereby reducing instances of contract fraud. Our framework supports sustainable development by promoting innovation, efficiency, and technological advancement for SDG 9. It also combats fraud, strengthens institutions, and fosters transparency and accountability for SDG 16 in the insurance sector. Future research will focus on leveraging prominent deep learning models globally and utilizing diverse datasets to refine and develop the most robust feasible model.
Insurance fraud, imbalanced dataset, detection, neural network, fraud prevention efforts, sustainable, SDGs. 2 TABLE OF CONTENTS 1. Thesis Approach and Contribution. Traditional model analysis.
Deep learning model. Formula of reLu, Batch Normalization and Drop out.2 Stacking learning model. Strategies to improve performance of model. Recursive Feature Elimination.
class_dict_weight and SMOTE. Hyperparameter tuning with RandomSearch. Materials and Methods. AutoMobile Claims Dataset and Challenges.
Our proposed method. Experiments on dataset. Experiments on Traditional Model. Experiments on Deep Learning Model.
Experiment on Stacking model. Conclusion & Future Works. 41 4 LIST OF FIGURES Figure 1: Illustration of the measurement of process times and quantities along the core stages of claims management.8 Figure 2: Illustration of how the ML-driven approach uses AI & replaces rule-based methods that require human interventions to detect fraud 9 Figure 3. Comparison of deep learning and traditional machine learning methods.
Source: (Chen et al. 10 Figure 4: The Support Vector Paradigm. 14 Figure 5: A graphical representation of Decision Trees. 16 Figure 6: The depiction of a 600-lree Random Forest.
The average of the 600 predictions is the final prediction of the Random Forest. Batch Normalization formula. 18 Figure 8: Stacking model process. 19 Figure 9: Workflow Diagram of Recursive Feature Elimination (RFE).
21 Figure 10: Synthetic Minority Over-sampling Technique. 22 Figure 11: RandomSearch layout. 22 Figure 12: The distribution of data in the Automobile Insurance Claims Dataset. 25 Figure 13: The correlation among variables in the Automobile Insurance Claims Dataset.
25 Figure 14: Segment the total amount of regular requests. 26 Figure 15: Features for fraud detection. 27 Figure 16: Information transformation. 28 Figure 17: Stacking model process.30 Figure 18: Process of experiment setup.31 Figure 19: Accuracy and Fl - score of traditional models.
34 Figure 20: The ROC curve of the traditional model. 35 Figure 21: Confusion Matrix for Random Forest and Decision Tree.36 Figure 22: Confusion Matrix for Stacking model. 37 5 LIST OF TABLES Table 1: Descriptive statistics for categorical variables in the dataset. 24 Table 2: The previous proposed models in previous analysis in this topic.
29 Table 3: A result for imbalanced methods in dataset. 33 Table 4: A result for experiment models in dataset. 38 6 LIST OF ABBREVIATIONS Ordinal Symbol Explain number abbreviation 1 AI Artificial intelligence 2 AUC Area under the ROC Curve 3 BG Gradient Boosting model 4 cv Cross-Validation 5 DTC Decision Tree Classifier 6 EDA Exploratory Data Analysis 7 KNC K Nearest Neighbour Classifier model 8 LDA Means Linear Discriminant Analysis 9 LSTM Long short-term memory 10 MCC Matthews Correlation Coefficient 11 ML Machine learning 12 NB Naive Bayes model 13 NN Neural Network 14 PCA Principal Component Analysis 15 ReLU Rectified linear unit 16 RF Random Forest 7 17 RFE Recursive Feature Elimination 18 RNN Recurrent neural network 19 ROC Receiver operating characteristic 20 SMOTE Synthetic Minority Over-sampling Technique 21 SRA Unsupervised Spectral Ranking for Anomaly 22 SVM Support Vector Machine 8 1. Problem define Automobile insurance compensates the insured for losses from vehicle incidents like accidents, theft, and damage.
It provides coverage for liability, vehicle damage, and medical expenses, with details varying by policy and location. As the number of vehicles on the road increases, the need for automobile insurance grows to protect against risks associated with driving. However, as the automobile insurance sector grows, it has also become a more frequent target for fraudulent activity. This includes false or exaggerated claims for financial gain, affecting insurers and honest policyholders alike (Kowshalya & Nandhini, 2018), (Derrig, 2002).
Fraudsters utilize various tactics, such as inflating service costs, staging accidents, or filing claims for incidents that never occurred, challenging the industry to develop robust fraud detection and prevention measures. Fraud detection is the process of identifying and preventing fraudulent activities. It involves using data mining, machine learning, and deep learning algorithms to automate the detection process and improve accuracy. By analyzing historical data and detecting patterns, these algorithms can recognize fraudulent behavior in real-time data.
In the context of automobile fraud detection, various methodologies have been used such as machine learning algorithms and LSTM RNN networks (Kabir, 2022). Furthermore, there are advanced statistical and machine learning models such as the "FraudBuster" framework, specifically created to detect potential fraud during the underwriting process by analyzing patterns in policy characteristics and loss ratios (Nagrecha et al. These endeavors highlight the significance of advanced analytics in combating insurance fraud and bolstering the integrity of the insurance sector. Additionally, data technologies have been adopted to identify instances of insurance fraud, while machine learning and AI are utilized in risk management and fraud detection (Bart, 2017b), (Aziz & Dowling, 2018).
Problem statement Insurance fraud involving fabricating evidence or making invalid claims can occur in various ways. One method is by fabricating insurance subject matter, where a non existent insurance subject matter is created to deceive the insurer (Zhen, Jing. Another way is by providing untruthful or incomplete information in insurance applications or claims, such as submitting a claim based on misleading circumstances or exaggerating a genuine claim (Bellas et al. These fraudulent practices not only 9 burden insurance companies financially but also result in higher premiums for honest policyholders.
To combat insurance fraud, deterrence and detection methods are employed, including specialized investigation units, statistical analysis of claims information, and fraud detection techniques (Feinman, 2016). Figure 1 below shows an overview of the process in common claims management for insurance companies Notification ) / Registration ỵ1 ■ Audit > Settlement ) Closing -------- 1--------- Occurrence Customer Claim is Finalized Evaluated Customer of claim informs systemically claim coverage adjustment receives payment insurer registered audit amount and closing note and segmented — Measurement of claims quantities and amounts Measurement period for claims cycle and work times Figure Ỉ: Illustration of the measurement of process times and quantities along the core stages of claims management. Source: (Mahlow & Wagner, 20Ì6) In automotive insurance, several techniques exist to identify fraudulent activities, which can occur either in a skilled or unskilled manner (Ajay et al. The goal is to detect multiple frauds involving behavior changes, which can be challenging for traditional machine learning approaches (Caruana & Grech, 2021).
The current traditional methods used in insurance fraud detection are usually financial evaluation methods. These methods suffer from performance degradation caused by imbalanced data and fail to consider the low rankness prior in the intrinsic samples when outliers emerge (Ali et al. Traditional models often struggle when handling imbalanced datasets, particularly in machine learning applications such as fraud detection. These models tend to produce suboptimal results for the minority class, which is vastly outnumbered by the majority class.
Various techniques have been explored to address this issue, including resampling methods. Oversampling techniques have been found to be particularly effective in improving model performance on the minority class, resulting in higher accuracy and lower false negatives (De Zarza et al. 10 TRADITIONAL RULE-BASED APPROACH Leads to Scammer Fraud Rules Detection MACHINE LEARNING APPROACH Leads to User Transaction The Engine Constantly Improving Detection (Al Platform) Figure 2: Illustration of how the ML-driven approach uses AI & replaces rule-based methods that recpiire human interventions to detect fraud. Source: (Sinha, 2023) Additionally, traditional statistics-based methods are found to be more cost-effective than machine-learning based methods in detecting automobile insurance fraud (Benedek et al.
Al-based fraud detection methods tested on a real database are also found to be less cost-effective than traditional statistical-econometric methods (Benedek et al. Therefore, the disadvantages of traditional methods in insurance fraud detection include performance degradation due to imbalanced data, failure to consider low rankness prior, and lower cost-effectiveness compared to traditional statistics-based methods and Al-based methods (Tongesai et al. Deep learning models arc generally faster than traditional model-based methods for data processing. Traditional model-based methods rely on mathematical formulations and can be sensitive to inaccuracies, leading to slower performance in complex or dynamic systems (Shen et al.
On the other hand, deep learning models learn their mapping from data and can operate in complex environments, resulting in faster processing. However, deep learning models typically require large training sets and computational resources, which can limit their applicability in certain scenarios (Kabir, Nagrecha. Therefore, while deep learning models are faster, they may not always be the most efficient choice for all data processing tasks. Hence, we have put forth a strategic plan to address the limitations of conventional models and integrate them with deep learning models in the subsequent manner.
Firstly, we will select the conventional models that yield the most optimal outcomes. In our 11 analysis of research articles, it was observed that Random Forest and Support Vector Machine, two traditional models, exhibit superior performance. Hence, secondly, our focus will be on enhancing the challenges faced by these two models. Our direction is to integrate contemporary deep learning models with the Stacking technique to enhance the accuracy of result prediction.
Comparison of deep learning and traditional machine learning methods. Source: (Chen et al. Related works In prior scholarly investigations, it has been ascertained that two distinct methodologies have been employed to identify instances of insurance fraud. The initial approach encompasses traditional models such as Logistic Regression, Support Vector Machine, Naive Bayes, and Random Forest.
These traditional models have been extensively utilized in the field to uncover instances of fraudulent activities within the insurance industry. In this paper, we aim to examine the Support Vector Machine (SVM) model by employing the kernel trick and radial basis function techniques (Rustam et al. Additionally, our research also encompasses an investigation into the Random Forest model, which will be analyzed based on the K-Nearest Neighbor classification mechanism (Li et al. Secondly, in order to embrace more contemporary approaches, we delve into the realm of deep learning models, employing methodologies like reLu, Batch Normalization, and Drop out.
Finally, we lay down the theoretical foundation of data processing techniques or feature selection to optimize the problem solving approach. Through this exploration, we aim to overcome the constraints imposed by conventional models and discover novel solutions. Thesis Approach and Contribution The insurance industry faces significant challenges, particularly in dealing with fraudulent claims, which result in substantial financial losses. Currently, conventional financial evaluation methods are predominantly used to assess fraud cases.
Recognizing the urgency and importance of addressing this issue, we have chosen this topic as an academic research theme to serve the property insurance sector and the insurance domain in general.