Automobile Insurance Claim Fraud Detection using Random Forest and ADASYN
0 likes•19 views
The document presents a study on detecting automobile insurance claim fraud using random forest and ADASYN. The researchers used the random forest classifier and the ADASYN data sampling technique to address the class imbalance in their dataset. They applied one-hot encoding to resolve issues with imbalanced data, trained the random forest model on the balanced data, and evaluated its performance. Experimental results found that the random forest model achieved over 97% accuracy, 94% recall, and 99.8% precision for fraud detection after applying ADASYN, outperforming other classifiers like SVM and Naive Bayes. Thus, the random forest model with ADASYN was effective for the task of automobile insurance fraud detection.
![International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 05 | May 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 104
Automobile Insurance Claim Fraud Detection using Random Forest and
ADASYN
Dhruvang Gondalia1, Omkar Gurav2, Ameya Joshi3, Aniruddha Joshi4, Prof. Sangeetha Selvan5
1,2,3,4UG Student, Dept. of Computer Engineering, Pillai College of Engineering, New Panvel, India
5Assistant Professor, Dept. of Computer Engineering, Pillai College of Engineering, New Panvel, India
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - With the increasing number offraudulentclaims
in the insurance industry, this issue needs to be contained. Car
insurance fraud is the most common compared to all other
types of fraudulent claims. Therefore, it is necessary to have a
system to detect and prevent such fraud, and it is necessary to
build a system to detect insurance fraud. Manyfraud detection
models are created using a variety of algorithms and
techniques. We used a random forest as a classifier and
ADASYN to balance the dataset. One HotEncodingwasusedto
resolve an issue of undesirable attributesduringbalancingthe
dataset. This application we created can be used by car
insurers to evaluate customer claims more quickly than other
traditional methods that involve manual tasks. Therefore, this
application helps find out if the claim is genuineorfraud while
the customer is claiming insurance. It is more accurate and
free of fraud than traditional methods. Other techniques such
as SVM can be used, but for this particular problem, Random
Forest seems ideal because it provides significantly better
accuracy than other techniques.
Key Words: ADASYN, SVM, Random Forest, Data
Sampling, Insurance Fraud, Fraud Detection, One Hot
Encoding
1. INTRODUCTION
Insurance fraud occurs when aninsuranceprovider,advisor,
adjuster, or consumer intentionally deceives in order to
obtain an illegal gain. There has been an increase in
fraudulent insurance claims in recent years, particularly in
the automobile insurance industry. Falsify insurance claim
information, exaggerate insurance claims to represent an
accident, or submit a claim form for damage or injury that
has never occurred by making a false claim for car theft.
That's all an example of a car insurance fraud. When
insurance companies use fraud detection systems, they not
only detect fraud but also save millions, if not billions, of
dollars that would otherwise be paid to the person who
made the fraudulent claim.
2. LITERATURE REVIEW
i. Detecting Fraudulent Insurance Claims UsingRandom
Forests and Synthetic Minority Oversampling
Technique: The author used SMOTE to balance the dataset
and used Random Forest for the prediction of the claim, So
SMOTE with random forest gives accuracy upto 94%. But it
can be improved by using other balancing techniques like
ADASYN which is grouped under over sampling technique
data balancing technique.
ii. Performance comparative study of machine learning
algorithms for automobile insurance fraud detection:
The author showed a study comparing ten of the most
frequently used machine learning algorithms for detecting
fraud in insurance claims. The study shows that theRandom
Forest algorithm has the best performance for insurance
fraud detection.
iii. Detecting Fraudulent Motor Insurance Claims Using
Support Vector Machines with Adaptive Synthetic
Sampling Method:They have used ADASYN forbalancing
the dataset where it tries to increase minority class samples
by adding similar entries in it. Base model used in this
project was SVM but the dataset used in it consists of only
1000 rows out of which 25% of the data consists of
fraudulent claim and rest were genuine claim.
iv. Automobile Insurance Fraud Detection using
Supervised Classifiers: The dataset used in this project is
not available on internet the dataset consists of 11 different
columns such as Gender of Policyholder, Police Report File
,Model of Car etc So for balancing thedatasettheauthorused
SMOTE to balance it and tested dataset with 3 different
classifier they are Multi-Layer perceptron,Decisiontree,and
Random forest, Author found that Random forest is best
technique for this problem statement.
v. Fraud Detection by Machine Learning: Here the author
discusses different types of credit card frauds. He proposed
the dataset should be in 1:1 ratio for fraud and genuine
cases. And he tested different machine learning algo such as
logistic regression, support vector machine, boosted trees,
random forest, and neural network etc. and found random
forest to be the best fit algorithm for his dataset.
3. Dataset and Parameters
The experimental dataset used in this study is provided by
the user Jwilda on kaggle[6]. The dataset has 15,420 rows
with 33 columns of data. Each row in the dataset has 33
attributes in total. Out of which, 32 are claim features that
will help to predict the last 1 variable, called the class label.
Here, FraudFound is our target variable which will containa](https://image.slidesharecdn.com/irjet-v9i523-220929062603-5078eae9/85/Automobile-Insurance-Claim-Fraud-Detection-using-Random-Forest-and-ADASYN-1-320.jpg)
![International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 09 Issue: 05 | May 2022 www.irjet.net p-ISSN: 2395-0072
© 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 107
We would like to express our gratitude to our Principal, Dr.
Sandeep Joshi, who gave us a golden opportunity to do this
wonderful project on the topic of ‘Automobile Insurance
Claim Fraud Detection using Random Forest and ADASYN’,
which has also helped us in doing a lot of research and
learning their implementation.
REFERENCES
[1] S. Harjai, S. K. Khatri and G. Singh, "DetectingFraudulent
Insurance Claims Using Random Forests and Synthetic
Minority Oversampling Technique," 2019 4th
International Conference on Information Systems and
Computer Networks (ISCON), 2019, pp. 123-128, doi:
10.1109/ISCON47742.2019.9036162.
[2] B. Itri, Y. Mohamed, Q. Mohammed and B. Omar,
"Performance comparative study of machine learning
algorithms for automobile insurance fraud detection,"
2019 Third International Conference on Intelligent
Computing in Data Sciences (ICDS), 2019, pp. 1-4, doi:
10.1109/ICDS47004.2019.8942277.
[3] C. Muranda, A. Ali and T. Shongwe, "Detecting
Fraudulent Motor Insurance Claims Using Support
Vector Machines with Adaptive Synthetic Sampling
Method," 2020 61st International Scientific Conference
on Information Technology and Management Scienceof
Riga Technical University (ITMS), 2020, pp. 1-5, doi:
10.1109/ITMS51158.2020.9259322.
[4] I. M. Nur Prasasti, A. Dhini and E. Laoh, "Automobile
Insurance Fraud DetectionusingSupervisedClassifiers,"
2020 International Workshop on Big Data and
Information Security (IWBIS), 2020, pp. 47-52, doi:
10.1109/IWBIS50925.2020.9255426.
[5] Y. Wei, Y. Qi, Q. Ma, Z. Liu, C. Shen and C. Fang, "Fraud
Detection by MachineLearning," 20202ndInternational
Conference on Machine Learning, Big Data andBusiness
Intelligence (MLBDBI), 2020, pp. 101-115, doi:
10.1109/MLBDBI51377.2020.00025.
[6] Jwilda, “Classifying Fraud by Decision Trees”, Kaggle,
Available:
https://www.kaggle.com/code/jwilda3/classifying-
fraud-by-decision-trees/data ,[Accessed30-Sept-2021]
BIOGRAPHIES
Dhruvang Gondalia
Omkar Gurav
Ameya Joshi
Aniruddha Joshi](https://image.slidesharecdn.com/irjet-v9i523-220929062603-5078eae9/85/Automobile-Insurance-Claim-Fraud-Detection-using-Random-Forest-and-ADASYN-4-320.jpg)
Automobile Insurance Claim Fraud Detection using Random Forest and ADASYN
- 1. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 09 Issue: 05 | May 2022 www.irjet.net p-ISSN: 2395-0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 104 Automobile Insurance Claim Fraud Detection using Random Forest and ADASYN Dhruvang Gondalia1, Omkar Gurav2, Ameya Joshi3, Aniruddha Joshi4, Prof. Sangeetha Selvan5 1,2,3,4UG Student, Dept. of Computer Engineering, Pillai College of Engineering, New Panvel, India 5Assistant Professor, Dept. of Computer Engineering, Pillai College of Engineering, New Panvel, India ---------------------------------------------------------------------***--------------------------------------------------------------------- Abstract - With the increasing number offraudulentclaims in the insurance industry, this issue needs to be contained. Car insurance fraud is the most common compared to all other types of fraudulent claims. Therefore, it is necessary to have a system to detect and prevent such fraud, and it is necessary to build a system to detect insurance fraud. Manyfraud detection models are created using a variety of algorithms and techniques. We used a random forest as a classifier and ADASYN to balance the dataset. One HotEncodingwasusedto resolve an issue of undesirable attributesduringbalancingthe dataset. This application we created can be used by car insurers to evaluate customer claims more quickly than other traditional methods that involve manual tasks. Therefore, this application helps find out if the claim is genuineorfraud while the customer is claiming insurance. It is more accurate and free of fraud than traditional methods. Other techniques such as SVM can be used, but for this particular problem, Random Forest seems ideal because it provides significantly better accuracy than other techniques. Key Words: ADASYN, SVM, Random Forest, Data Sampling, Insurance Fraud, Fraud Detection, One Hot Encoding 1. INTRODUCTION Insurance fraud occurs when aninsuranceprovider,advisor, adjuster, or consumer intentionally deceives in order to obtain an illegal gain. There has been an increase in fraudulent insurance claims in recent years, particularly in the automobile insurance industry. Falsify insurance claim information, exaggerate insurance claims to represent an accident, or submit a claim form for damage or injury that has never occurred by making a false claim for car theft. That's all an example of a car insurance fraud. When insurance companies use fraud detection systems, they not only detect fraud but also save millions, if not billions, of dollars that would otherwise be paid to the person who made the fraudulent claim. 2. LITERATURE REVIEW i. Detecting Fraudulent Insurance Claims UsingRandom Forests and Synthetic Minority Oversampling Technique: The author used SMOTE to balance the dataset and used Random Forest for the prediction of the claim, So SMOTE with random forest gives accuracy upto 94%. But it can be improved by using other balancing techniques like ADASYN which is grouped under over sampling technique data balancing technique. ii. Performance comparative study of machine learning algorithms for automobile insurance fraud detection: The author showed a study comparing ten of the most frequently used machine learning algorithms for detecting fraud in insurance claims. The study shows that theRandom Forest algorithm has the best performance for insurance fraud detection. iii. Detecting Fraudulent Motor Insurance Claims Using Support Vector Machines with Adaptive Synthetic Sampling Method:They have used ADASYN forbalancing the dataset where it tries to increase minority class samples by adding similar entries in it. Base model used in this project was SVM but the dataset used in it consists of only 1000 rows out of which 25% of the data consists of fraudulent claim and rest were genuine claim. iv. Automobile Insurance Fraud Detection using Supervised Classifiers: The dataset used in this project is not available on internet the dataset consists of 11 different columns such as Gender of Policyholder, Police Report File ,Model of Car etc So for balancing thedatasettheauthorused SMOTE to balance it and tested dataset with 3 different classifier they are Multi-Layer perceptron,Decisiontree,and Random forest, Author found that Random forest is best technique for this problem statement. v. Fraud Detection by Machine Learning: Here the author discusses different types of credit card frauds. He proposed the dataset should be in 1:1 ratio for fraud and genuine cases. And he tested different machine learning algo such as logistic regression, support vector machine, boosted trees, random forest, and neural network etc. and found random forest to be the best fit algorithm for his dataset. 3. Dataset and Parameters The experimental dataset used in this study is provided by the user Jwilda on kaggle[6]. The dataset has 15,420 rows with 33 columns of data. Each row in the dataset has 33 attributes in total. Out of which, 32 are claim features that will help to predict the last 1 variable, called the class label. Here, FraudFound is our target variable which will containa
- 2. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 09 Issue: 05 | May 2022 www.irjet.net p-ISSN: 2395-0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 105 value, either ‘1’ or ‘0’. This variable represents whether the claim is genuine or fraud. ‘1’ would mean the claim is fraud and value ‘0’ represents a genuine claim. Here 25 out of 32 claim features are categorical and remaining 7 features are numerical. Out of 15,420 rows, 14,497 rows consist of genuine claim data and the rest 923 rows consist of fraudulent claims. So the number of genuineclaimsisalmost 15 times more than the number of fraudulent claims. So the number of fraudulent claims is negligible compared to genuine claims. This creates a class imbalance, which will lead to a biased prediction model. In order to tackle this problem, data balancing is required. 4.1 System Architecture The system architecture is given in Figure 1. Each block is described in this Section. Fig.1 Proposed system architecture A. Data preprocessing: Here we checked our data for any missing values, redundant data, duplicates or null values so we removed those rows from the training dataset. Also, we transformed the categorical data into numeric data by using label encoding and a few columns with One Hot Encoding. Along with that, few columns consistofa rangevaluelike 10- 20 so we replaced these values with mean of their extreme values. We also maintained a dictionary to get back categorical value from label encoded value. B. Feature Selection: Based on the literature survey we made we have selected an important column from the dataset. We also removed some of the unwanted columns like Policy number which consists of random valuesforeach insurance claim and does not affect policy claims. C. Applying One Hot encoding: One hot encoding is one of the techniques to represent a categorical feature. Here we set a new binary variable for each unique value in a categorical feature. It is oneofthemostpreferredtechniques when it comes to training categorical data. But its disadvantage is that the number of columns is equal to the number of unique values in the column of the categorical dataset. We have used One Hot encoding because directly using ADASYN created undesirable values for some attributes for e.g., Age is a whole number value which was a fractional value in the dataset generated by ADASYN. D. Data Balancing using ADASYN: To train a model forsuch classification where the number element in one class is less than the other class in such a situation our model will make biased predictions where we see our model to be more tend towards the majority class. So, ADASYN is one of the data balancing techniques which tries to increase the number of minority class samples. E. One hot encoding to Label Encoding: Once we generated random samples in one hot encoding but the issue here is that the number of columns increased from 33 to 105 which is a very tedious task to handle such a huge data column. So we have converted it back to categorical data so wewill geta balanced dataset with valid inputs for the model. But again we cannot provide categorical data to train the model so we converted these categorical features to numeric labels. F. Data Splitting: For training the model we will split the dataset in two parts for example 25% of data for testing and remaining 75% of data for training. G. Training: We have used a few machine learning algorithms, which trains the data set aside for training the dataset into a model which will classify any new input case as fraud or not fraud. The algorithms which we used are Support Vector Machine (SVM), Naive Bayes, AdaBoost and Random Forest. H. Testing: Remaining of splitted data is used for testing the model. Output of this will help us to evaluate our model using different evaluation metrics. I. Trained Model: Once we are done with Testing ourmodel. So, we finally create our model and test with different train- test split ratio and different randomness in our dataset. And we find the best configuration model for our problem statement. Now this model is ready to give us a prediction whether a new insurance claim is fraud or genuine. 5. Performance Analysis A. Evaluation Criteria: There are different evaluation metrics for evaluation of our model, few of the popular metrics are accuracy, precision and recall. These metrics are calculated using a confusion matrix which is prepared in the Testing phase of ourproject. A confusion matrix consists of 4 different values: True positive, True Negative, False Positive and False Negative. These are calculated as the number of cases classified as genuine and they are actually genuine, these claims are called as True Positive. Similarly, if a claim is fraudulent and it is classified as fraudulent then it is called as true negative. These are the two values which show both positive and negative classes which are correctly classified.
- 3. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 09 Issue: 05 | May 2022 www.irjet.net p-ISSN: 2395-0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 106 Fig.2 Confusion matrix Based on the above matrix one can evaluate his model by finding differentvaluessuchasaccuracy,precisionandrecall as Fig.3 Formula for Recall The above equation can be explained by saying that, from all the positive classes, what percentage correct we predicted. Fig.4 Formula for Precision The above formula can be explained by showing how many of all the classes that were predicted to be positive are actually positive. Accuracy= TP+TN/(TP+TN+FP+FN) Accuracy is calculated as a percentage of how many entries we correctly classified as correct to the total number of entries. B. Experimental Results After oversampling the minority class in the dataset using ADASYN the number of rows wereincreasedto28,628out of which 14410 are of fraud claims and 14208 are genuine claims. Fig.6 Confusion matrix for Random Forest with ADASYN For testing our model, we gave 7155 rows which is 25% of the total balanced dataset. Out of these we got True Positive: 3390, True Negatives: 3559, False Positives: 6 these are the claims which are classified as fraud but labeled as genuine False Negatives: 200 thesearetheclaimswhichareclassified as genuine but they are labeled as fraud. Table 1: Comparison of various classifiers on balanced dataset Performance Metrics (in%) Support Vector Machine (SVM) Naive Bayes AdaBoost Random Forest Accuracy 62.4 88.5 95.8 97.1 Sensitivity (or recall value) 84.5 91.1 92.7 94.4 Precision 58.7 86.5 98 99.8 In the above table results of various classifiers are given. Random Forest has performed better than other classifiers in all three metrics Accuracy, Sensitivity and Precision. SVM did not perform much well for this dataset. AdaBoost and Naive Bayes gave pretty good accuracy but not better than random forest. ACKNOWLEDGMENT We would like to extend our deepest gratitudetoourProject Guide, Prof. Sangeetha Selvan, who guided us and provided us with her valuable knowledge and suggestions on this project and helped us improveourprojectbeyondourlimits. We would also like to express our heartfelt thanks to our Head of Department, Dr. Sharvari S. Govilkar, for providing us with a platform where we can try to work on developing projects and demonstrate the practical applications of our academic curriculum.
- 4. International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 09 Issue: 05 | May 2022 www.irjet.net p-ISSN: 2395-0072 © 2022, IRJET | Impact Factor value: 7.529 | ISO 9001:2008 Certified Journal | Page 107 We would like to express our gratitude to our Principal, Dr. Sandeep Joshi, who gave us a golden opportunity to do this wonderful project on the topic of ‘Automobile Insurance Claim Fraud Detection using Random Forest and ADASYN’, which has also helped us in doing a lot of research and learning their implementation. REFERENCES [1] S. Harjai, S. K. Khatri and G. Singh, "DetectingFraudulent Insurance Claims Using Random Forests and Synthetic Minority Oversampling Technique," 2019 4th International Conference on Information Systems and Computer Networks (ISCON), 2019, pp. 123-128, doi: 10.1109/ISCON47742.2019.9036162. [2] B. Itri, Y. Mohamed, Q. Mohammed and B. Omar, "Performance comparative study of machine learning algorithms for automobile insurance fraud detection," 2019 Third International Conference on Intelligent Computing in Data Sciences (ICDS), 2019, pp. 1-4, doi: 10.1109/ICDS47004.2019.8942277. [3] C. Muranda, A. Ali and T. Shongwe, "Detecting Fraudulent Motor Insurance Claims Using Support Vector Machines with Adaptive Synthetic Sampling Method," 2020 61st International Scientific Conference on Information Technology and Management Scienceof Riga Technical University (ITMS), 2020, pp. 1-5, doi: 10.1109/ITMS51158.2020.9259322. [4] I. M. Nur Prasasti, A. Dhini and E. Laoh, "Automobile Insurance Fraud DetectionusingSupervisedClassifiers," 2020 International Workshop on Big Data and Information Security (IWBIS), 2020, pp. 47-52, doi: 10.1109/IWBIS50925.2020.9255426. [5] Y. Wei, Y. Qi, Q. Ma, Z. Liu, C. Shen and C. Fang, "Fraud Detection by MachineLearning," 20202ndInternational Conference on Machine Learning, Big Data andBusiness Intelligence (MLBDBI), 2020, pp. 101-115, doi: 10.1109/MLBDBI51377.2020.00025. [6] Jwilda, “Classifying Fraud by Decision Trees”, Kaggle, Available: https://www.kaggle.com/code/jwilda3/classifying- fraud-by-decision-trees/data ,[Accessed30-Sept-2021] BIOGRAPHIES Dhruvang Gondalia Omkar Gurav Ameya Joshi Aniruddha Joshi