A Survey on Fraud Analytics Using Predictive

Model in Insurance Claims 1K. Ulaga Priya and

2S. Pushpa

1Dept of Computer Science and Engineering,

St.Peters University.

ulagapriya@gmail.com 2Dept of Computer Science and Engineering,

St.Peters University.

pushpasangar96@gmail.com

Abstract

Insurance Industry is a rapidly growing fast industry in terms of large

amount of data. The most critical issue in insurance industry is fraudulent

claims. Fraud is nothing but wrongful or criminal trick planned to result in

financial or personal gains. As the size of data increases, the traditional

approach will not work and it will be tedious job to identify the fraudulent

claims. Moreover, new types of claim will emerge and hence it will be

difficult to predict the fraudulent claims. This paper depicts an overview of

Fraud analytics, prediction, and Data Science algorithms based predictions

in insurance industry.

International Journal of Pure and Applied MathematicsVolume 114 No. 7 2017, 755-767ISSN: 1311-8080 (printed version); ISSN: 1314-3395 (on-line version)url: http://www.ijpam.euSpecial Issue ijpam.eu

755

1. Introduction

Fraud analytics is a type of data analytics where data analysis is done on the

fraudulent behaviour. There are several domains where fraud may happen like

Credit card fraud, telecommunication fraud, Insurance fraud, Healthcare fraud,

tax evasion etc. Credit card fraud is one of the fraud types which is surveyed

widely in the domain of fraud detection.[34],[35,[36].Due to the popular mode

of payment transaction, both online and offline, the fraud associated with it is

also increasing. There are multiple techniques to detect credit card fraud like

Neural Network [10-11],Group Method of Data Handling [4-5], Bagging[6].

Some other popular models of credit card fraud detection are Hidden Markov

Model [2-3], Bayesian learning [7-9], K-means Clustering [1].The credit card

fraud was categorised [35] as two categories namely behavioural frauds and

Application frauds. Application frauds happen whenever fraudsters[33] acquire

new cards by providing false data to issuing companies[33]. Behavioural frauds

include four types: mail theft, fake cards, stolen/lost cards. Several algorithms

[43] in credit card fraud prediction were compared and derived that Bagging

ensemble classifier is the best method.

Telecommunication fraud is rapidly increasing due to the growth of recent

technology and global communication which results in considerable losses in

business. There are two categories of telecommunication fraud: subscription

fraud and super imposed fraud. Subscription fraud is nothing but claiming false

identity for getting service and elude payment. Superimposed fraud happens

whenever the service is used without having relevant rights and is usually

detected by the appearance of 'phantom' call on a bill. Various techniques used

in telecommunication fraud detection[12] are Neural Networks, Visualization

Methods and Rule-based Approach.

Insurance fraud is defined [37] as fraud in the insurance industry as perceptively

creating a fabricated claim, bloating a claim or adding further items to a claim,

or being in any way deceitful with the intention of getting more than legitimate

privilege. The insurance fraud types include exaggerated claims, fabricated

medical history, post-dated policies, faked damage etc. [30] This emphasize the

different types of fraud in health insurance sector. There are different techniques

for health insurance fraud detection[22]. This paper concentrates on Insurance

Fraud and its data analytics. The National Healthcare Anti-Fraud Association

(NHCAA) evaluated the health care claims and announced that 10 percent of

health care claims contain some element of fraud [38][39]. Insurance protects

the customer from monetary loss. Insurance Policy is a legal agreement between

the Policy holder and insurance [23] company which specifies the claim amount

which the Policy Holder needs to pay. Insurance claim is nothing but, the policy

holder request the claim amount from the insurance company based on the

insurance policy. Insurance domain can be categorised as (i)Health Insurance.

(ii) Travel Insurance (iii) Auto insurance (iv) Life insurance

International Journal of Pure and Applied Mathematics Special Issue

756

The section2 describes about Bigdata analytics in identifying fraudulent claims.

Section3 describes one of the Bigdata Analytics type which is predictive

Analytics. This section also discusses the various types of Machine learning

algorithms, Section4 discusses the merits and demerits of the algorithm, it also

explains the Fraud analytics process model. Section5 discusses the performance

benchmark of different types of fraud. Section6 depicts the conclusion and

Section7 holds the references.

2. Big Data Analytics

Fraud detection in insurance is a potential area in insurance where big data

plays a major role. However, many insurers remain unknown about the power

of data analytics. According to the survey conducted almost 80% of insurer is

unaware about the power of Big Data Analytics. Let us examine a few data

analytic models that can help insurers strengthen their fraud detection

capabilities.

i) Descriptive – Analysing the data on what was already happened. Generally,

reports were generated with past data and analysis is done on that data. For

example, to identify the sales distribution that has happened in previous year.

ii) Diagnostic – Based on the previous data, data analysis will be done on why it

is has happened. Identifying and analysing the reason for poor sales in the

previous year is an example of diagnostic data analysis.

iii) Predictive–This type of analysis will suggest[27] what will happen in the

future. It predicts the futuristic scenario based on past historical data. For

example, identifying the area that is likely to perform better sales in the current

year based on past data.

iv)Prescriptive – This type of analysis will suggest what action should be taken.

Basically, how we can make it happen. It gives recommendation on what needs

to be done. For example, how to achieve the best outcome in sales, and strategy

to retain key customers.

3. Predictive Analytics

This paper discusses on predictive analytics and the techniques used for

prediction. Supervised learning and Unsupervised learning are the [28]

techniques used for predictive analytics. Supervised Learning will have a target

International Journal of Pure and Applied Mathematics Special Issue

757

variable. Target variable is the output that is predicted using other relevant

features. Unsupervised Learning does not have a target variable. Following

supervised learning techniques are used for predicting fraudulent claims since it

has a target variable.

Decision tree

Random Forest

Support Vector Machines

Neural Network

XGBoost

These techniques are used in solving data analytics problems.

Decision Tree

Decision tree gives a visualisation view in the form of graph. The sample set is

divided into subset of trees which represent choices and their results. Each node

of a tree represents a choice and the edges represent the decision. The sample

dataset is categorised into training dataset and test dataset. A model is created

with training dataset which gives the prediction accuracy. This model is applied

on the test dataset and the accuracy of prediction is validated. For each predictor

variable, this model can be used to decide on the category(Yes/No, Spam/not

spam-) of the data.Decision tree can deal with continuous data through various

method of decision tree like ID3 method and C4.5.

Decision tree is used in various fraud detection and prediction applications.

Some of the fraudulent problem areas where decision tree is used are credit card

fraud, Energy fraud etc. Credit card fraud detection [40] uses a cost sensitive

decision tree approach. Decision tree is also effectively used in Energy Fraud

detection. This technique is widely used for classification and regression. M5P

Decision tree is used for energy fraud detection which is a modified version of

Quinlan’s [12] M5 algorithm. Following is the general algorithm. Input: Training dataset

Output :To create a decision Tree.

Step 1: Identify the best attribute of the dataset which need to be placed at the

root of the tree.

Step 2: Divide the training set into subsets. Each subset should contain data

with the same value such that each subset is created for an attribute.

Step 3: Till you find leaf nodes step1 and step2 is to be repeated on each

subset in all the branches of the tree.

International Journal of Pure and Applied Mathematics Special Issue

758

Random Forest

In the random forest technique, multiple decision trees are created. A random

subset of the training data is used to create a single decision tree. [16] The

common result of each random subset is taken as the final tree output. A new

study is fed into all the trees and majority vote for each classification was taken

in this model. Missing values and outliers are taken care in random forest

model.

The predictive algorithm which uses this technique will try to imitate the

relationship between input and output variable. This algorithm provides

excellent accuracy and it runs very effectively on large datasets. This

algorithm[14] is widely used for large number of input. Moreover, it has

methods for maintaining balance for the unbalanced datasets

It is identified that for the aggregated model random forest gives better results

than Naïve Bayes. Where as in the personalized models Naïve Bayes gives

better results. In online shopping [15] when large number of discounts are

announced, it paves way for unusual activities in purchasing products and

services. This paper uses random forest algorithm to detect faults using R

language. Prediction can be done using Random Forest technique to identify

customer’s preference regarding the choice of insurance policy options. [12].

Following is the algorithm: Input: Training dataset

Output: To create “n” of Trees

Step 1: Randomly select “k” features from total “m” features Where k << m

Step 2: Among the “k” features, calculate the node “d” using the best split point.

Step3: Split the node into daughter nodes using the best split.

Step4: Repeat 1 to 3 steps until “l” number of nodes has been reached.

Step 5: Build forest by repeating steps 1 to 4 for “n” number times to create “n” number of

trees.

Neural Networks

The fundamental element of computation in neural network is the neuron which

is also called as node or unit. The input from other nodes is computed and

produces an output. Basically, it converts the input from multiple sources to

output. Whereas in human brains has a distinct feature of creating transient

states through neurons in between sensory organ and brain which is the decision

International Journal of Pure and Applied Mathematics Special Issue

759

taking unit.

To detect and predict the risk of fraudulent financial reporting, a Multilayer

Perceptron (MLP) [17] Artificial Neural Network model was proposed.

Weatherford suggests, artificial immune systems, recurrent neural networks,

back propagation neural networks for fraud detection. A neural network

approach is identified to detect management fraud. The management fraud is

detected [18] using the Adaptive Logic Network and generalized adaptive

neural network.[42] A three-layer was used with feed-forward Radial Basis

Function (RBF) neural network which will produce in every two hours for new

credit card transactions. This also propose fuzzy neural networks on parallel

machines which rises the rule production for customer-specific credit card fraud

detection. Neural Network gave better results for prediction when compared to

Logistic Regression ad Decision Tree[18]. A case study was done with 5

strategies to audit the auto insurance claims[9].

Input: Training dataset

Output: To create data model.

Step 1:Assign random weights to all the linkages to start the algorithm

Step 2: Using the inputs and the (input-hidden node) linkages find the activation rate

of hidden nodes

Step 3: Using the activation rate of hidden nodes and linkages to output, find the

activation rate of output nodes

Step 4: Find the error rate at the output node and recalibrate all linkages between

hidden nodes and output nodes

Step 5: Using the weights and error found at the output node, cascade down the error

to

hidden nodes

Step 6:Recalibrate the weights between hidden node and the input nodes repeat the

process till the convergence criterion is met

Step 7: Using the final linkage weights score the activation rate of the output nodes

XGBoost

XGBoost is a short form for Extreme Gradient Boosting. Boosting is a

sequential process. Multiple trees are created and the information of the first

tree is fed as input to the second tree so that it improves the prediction in

subsequent iterations. Basically it is a additive tree model where it add new

trees that complement the already built ones. XGBoost handles missing values

and it works only for numeric data.

Support Vector Machine

Support Vector Machines (SVM) is also a supervised learning algorithm used

for regression and classification problems. In general, it creates a hyper plane in

International Journal of Pure and Applied Mathematics Special Issue

760

n dimensional space to classify the data based on target class. The SVM

separates into different classes through a hyperplane or multiple hyperplane.

The hyperplane separates the data points and sometimes it is difficult to separate

the data point through a single hyperplane. The distance between the data point

and hyperplane represents a margin.

This enables to perform classification or regression also. Since it has many

features SVM becomes a promising technique in prediction. [25]Basically,

SVM works on the principle that data points are segregated through

hyperplanes. This subsequently maximizes the distance between data points,

and the hyper plane is constructed with the help of support vectors. A Turkish

insurance company database [19] was taken for research. SVM technique was

applied to this data. SVM is basically a classification technique that identifies

each record as anomalous or normal record. Subsequently every record is

checked with margin and based on that the record is treated as normal or

anomalous. SVM is a kernel based [19] algorithm where kernel transmutes the

input data points to a high-dimensional space so that the problem is solved. [25]

There are different applications which detect fraud through SVM. The top

management fraud is detected using SVM, to create the Fraudulent Financial

statement. [20]

4. Discussion

A comparative study is done on the Supervised Technique. Each technique has

its own merits and demerits. Based on the application area and data technique

can be chosen and analytics can be done on that. The merits and demerits are

discussed below as follows:

Fraud Analytics Process Model

As a first step the business problem must be clearly identified. Next step is to

identify the data source which is a very important task in data analysis model.

[29] Then subsequently all the data is gathered in one single area which could

be a data mart or data warehouse. Then the data is cleaned up re, inconsistent,

International Journal of Pure and Applied Mathematics Special Issue

761

missing and duplicate values are removed. Additional data transformation is

done like data type conversion etc. In analytics phase, data model is built and

data is analysed with the newly created model. Once data analytics is done, this

will be examined by functional experts.[30] During the analytics phase, the

requirement of additional data may be identified. This triggers the need for

another round of data cleaning and transformation. The Pre-processing phase is

most time consuming[31].

5. Performance Benchmark for Different Types of Fraud

The following Scatter plot shows [39] unique fraud types which were discussed

and published in various fraud detection papers These were some of the

common fraud types highlighted in the Scatter Plot.

Different Types of Fraud

The following table provides references with performance metric of different

Fraud Types. For better comparison of different types of fraud the area under

Receiver operating characteristic(ROC) curve are only included, Reference Fraud Type Dataset Size Used PERCENTAGE

OFClass Distribution Performance Metric-Area under the curve

Ortega Figuerora

et al,(2006)

Medical Insurance 8,819 5% AUC 74%

Subelj Furlan et al.(2011)

Automobile Insurance fraud

3.451 1.3% AUC 71%-92%

Battacharyya Jha

et al(2011)

Credit card fraud 50 million transactions on about

1 million credit cards from a

single country

0.005% AUC 90.8% -95.3%

Whitrow , Hand

et al.(2009)

Credit Card Fraud 33,000 -36,000 activity records 0.1% Gini 85%

(~AUC=92,5%)

Van Vlasselaer Bravo et

al.(2015)

Credit Card Fraud 3.3 million transactions <1% > AUC 98.6%

Dongshan and

Girolami(2017)

Telecommunication

Fraud

809,395 calls from 1,067

accounts

0.024 AUC 99.5%

Van Vlasselaer

Meskens et al

(2013)

Social Security Fraud 2000 observations 1% AUC 80%-85%

International Journal of Pure and Applied Mathematics Special Issue

762

6. Conclusion

Like Insurance fraud detection, several fraudulent behaviours are available like

Intrusion detection fraud, credit card fraud, telecommunication fraud etc. It is

prominent that health insurance[21] fraud is viable since it brings heavy loss

overall. By integrating big data technology these claims can be predicted for

large volume of data as well as different variety of data .

References

[1] Srivastava A., Kundu A., Sural S., Majumdar A., Credit Card Fraud Detection Using Hidden Markov Model, IEEE Transactions On Dependable And Secure Computing 5(1) (2008), 37-48.

[2] Bhusari V., Patil S., Study of Hidden Markov Model in Credit Card Fraudulent Detection, International Journal of Computer Applications 20(5) (2011).

[3] Ivakhnenko A.G., The group method of data handling in prediction problems, Sov Autom Control 9(6) (1976), 21–30.

[4] Mueller J.A., Lemke F., Self-organising data mining: an intelligent approach to extract knowledge from data, Script Software International, Berlin (2009).

[5] Singh S.P., Shukla S.S.P., Rakesh N., Tyagi V., Problem Reduction In Online Payment System Using Hybrid Model, International Journal of Managing Information Technology 3(3) (2011).

[6] Zreapoor M., Shamsolmoali P., Application of Credit Card Fraud Detection: Based on Bagging Ensemble Classifier, International Conference on Computer, Communication and Convergence (2015).

[7] Benson Edwin Raj S., Annie Portia A., Analysis on Credit Card Fraud Detection Methods, International Conference on Computer, Communication and Electrical Technology (2011).

[8] Panigrahi S., Kundu A., Sural S., Majumdar A.K., Credit card fraud detection: A fusion approach using Dempster-Shafer theory and Bayesian learning, Special Issue on Information Fusion in Computer Security 10(4) (2009), 354-363

[9] Chang R.I., Lai L.B., Su W.D., Wang J.C., Kouh, J.S., Intrusion Detection by Backpropagation Neural Networks with Sample-Query and Attribute-Query, Research India Publications (2006).

[10] Patidar R., Sharma L., Credit Card Fraud Detection Using Neural Network, International Journal of Soft Computing and Engineering 1 (2011).

International Journal of Pure and Applied Mathematics Special Issue

763

[11] Guo T., Li G.Y., Neural Data Mining For Credit Card Fraud detection, Proceedings of the Seventh International Conference on Machine Learning and Cybernetics (2006).

[12] Lata L.N., Koushika I.A., Hasan S.S., A Comprehensive Survey of Fraud Detection Techniques, International Journal of Applied Information Systems 10(2) (2015).

[13] Quinlan J., Learning with continuous classes, 5th Australian joint conference on artificial intelligence 92 (1992).

[14] Alshamsi A.S., Predicting car insurance policies using random forest, 10th International Conference on Innovations in Information Technology (2014), 128-132.

[15] Viaenea S., Auto claim fraud detection using Bayesian learning neural networks, Elsevier (2005).

[16] Eesha Goel, Abhilasha, Ankit Agarwal, Fraud Detection Using Random Forest Algorithm, International Journal of Computer Science Engineering 5(05) (2016).

[17] Salama A.S., Omar A.A., A Back Propagation Artificial Neural Network based Model for Detecting and Predicting Fraudulent Financial Reporting, International Journal of Computer Applications 106(2) (2014).

[18] Fanning K., Cogger K.O., Srivastava R., Detection of management fraud: A neural network approach. Intelligent Systems in Accounting, Finance and Management 4(2) (1995), 113-126.

[19] Kirlidog M., Asuk C., A fraud detection approach with data mining in health insurance, Procedia-Social and Behavioral Sciences 62 (2012), 989-994.

[20] Pai P.F., A support vector machine-based model for detecting top management fraud, Knowledge-Based Systems 24 (2011), 314–321.

[21] Rawte V., Anuradha G., Fraud Detection in Health Insurance using Data Mining Techniques, Communication, Information & Computing Technology (2015).

[22] Peng Y., Kou G., Sabatka A., Chen Z., Khazanchi D., Shi Y., Application of clustering methods to health insurance fraud detection, International Conference on Service Systems and Service Management 1 (2006), 116-120.

[23] Thornton D., Mueller R.M., Schoutsen P., van Hillegersberg J., Predicting healthcare fraud in medicaid: a multidimensional data model and analysis techniques for fraud detection, Procedia technology 9 (2013), 1252-1264.

International Journal of Pure and Applied Mathematics Special Issue

764

[24] Lin F., Yeh C.C., Lee M.Y., The use of hybrid manifold learning and support vector machines in the prediction of business failure, Knowl. based Syst. (2010), 95–101.

[25] Tang X., Zhuang L., Cai J., Li C., Multi-fault classification based on support vector machine trained by chaos particle swarm optimization, Knowl. based Syst. 23(5) (2010), 486–490.

[26] Wan S., Lei, T.C., A knowledge-based decision support system to analyze the debris-flow problems at Chen-Yu-Lan River, Taiwan, Knowledge-Based Systems 22(8) (2009), 580-588.

[27] Hafiz K.T., Aghili S., Zavarsky P., The use of predictive analytics technology to detect credit card fraud in Canada, 11th Iberian Conference on Information Systems and Technologies (2016), 1-6.

[28] Alfred R., The rise of machine learning for big data analytics, 2nd International Conference on Science in Information Technology (2016).

[29] Banarescu A., Detecting and Preventing Fraud with Data Analytics, Elsevier (2015).

[30] Thornton D., Brinkhuis M., Amrit C., Aly R., Categorizing and Describing the Types of Fraud in Healthcare, Procedia Computer Science 64 (2015), 713-720.

[31] Lata L.N., Koushika I.A., Hasan S.S., A Comprehensive Survey of Fraud Detection Techniques, International Journal of Applied Information Systems (2015).

[32] Dal Pozzolo A., Caelen O., Le Borgne Y.A., Waterschoot S., Bontempi G., Learned lessons in credit card fraud detection from a practitioner perspective, Expert systems with applications 41(10) (2014), 4915-4928.

[33] Mahmoudi N., Duman E., Detecting credit card fraud by Modified Fisher Discriminant Analysis, Expert Systems with Applications 42(5) (2014), 2510-2516.

[34] Chan P.K., Fan W., Prodromidis A.L., Stolfo S.J., Distributed data mining in credit card fraud detection, IEEE Intelligent Systems and Their Applications 14(6) (1999), 67-74.

[35] Bolton R., Hand D., Unsupervised Profiling Methods for Fraud Detection, Credit Scoring and Credit Control VII (2001).

[36] Brause R.W., Langsdorf T.S., Hepp H.M., Credit card fraud detection by adaptive neural data mining, Internal Report 7/99 (J. W. Goethe-University, Computer Science Department, Frankfurt, Germany) (1999).

International Journal of Pure and Applied Mathematics Special Issue

765

[37] Gill K.M., Woolley, K.A., Gill M., Insurance fraud: The business as a victim. In M. Gill (Ed.), Crime at work, Leicester: Perpetuity Press (1994).

[38] Frieden J., Fraud Squads Target Suspect Claims, Business & Health 9(4) (1991), 21-33.

[39] Guzzi R., Furious About Fraud, Best's Review-Life/Health Insurance Edition (1989).

[40] Sahin Y., Bulkan S., Duman E., A cost-sensitive decision tree approach for fraud detection, Expert Systems with Applications 40(15) (2013), 5916-5923.

[41] Vapnik V.N., Estimation of Dependences Based on Empirical Data, Addendum 1, New York: Springer-Verlag (1982).

[42] Reilly D.L., Cooper L.N., Elbaum C., A neural model for category learning, Biological Cybernetics 45(1) (1982), 35-41.

[43] Zareapoor M., Application of Credit Card Fraud Detection: Based on Bagging classifier, Elsevier (2015).

International Journal of Pure and Applied Mathematics Special Issue

766

767

768