Insurance fraud detection with unsupervised deep learning

• Published date: 01 September 2021
• Author: Chamal Gomes,Zhuo Jin,Hailiang Yang
• Date: 01 September 2021
• DOI: http://doi.org/10.1111/jori.12359

J Risk Insur. 2021;88:591–624. wileyonlinelibrary.com/journal/jori

|

591

Received: 15 May 2020

|

Revised: 18 May 2021

|

Accepted: 26 June 2021

DOI: 10.1111/jori.12359

ORIGINAL ARTICLE

Insurance fraud detection with unsupervised

deep learning

Chamal Gomes

1

|Zhuo Jin

1

|Hailiang Yang

2

1

Centre for Actuarial Studies,

Department of Economics, The

University of Melbourne, Parkville,

Victoria, Australia

2

Department of Statistics and Actuarial

Science, The University of Hong Kong,

Pokfulam, Hong Kong

Correspondence

Zhuo Jin, Centre for Actuarial Studies,

Department of Economics, The

University of Melbourne, Parkville,

VIC 3010, Australia.

Email: zjin@unimelb.edu.au

Funding information

Research Grants Council, University

Grants Committee,

Grant/Award Number: 17304921

Abstract

The objective of this paper is to propose a novel deep

learning methodology to gain pragmatic insights into

the behavior of an insured person using unsupervised

variable importance. It lays the groundwork for un-

derstanding how insights can be gained into the frau-

dulent behavior of an insured person with minimum

effort. Starting with a preliminary investigation of the

limitations of the existing fraud detection models, we

propose a new variable importance methodology in-

corporated with two prominent unsupervised deep

learning models, namely, the autoencoder and the

variational autoencoder. Each model's dynamics is

discussed to inform the reader on how models can be

adapted for fraud detection and how results can be

perceived appropriately. Both qualitative and quanti-

tative performance evaluations are conducted, al-

though a greater emphasis is placed on qualitative

evaluation. To broaden the scope of reference of fraud

detection setting, various metrics are used in the qua-

litative evaluation.

KEYWORDS

autoencoder, insurance fraud detection, unsupervised deep

learning, variable importance, variational autoencoder

© 2021 American Risk and Insurance Association

1|INTRODUCTION

Insurance fraud is defined by Gill et al. (2005)as“knowingly making a fictitious claim, inflating

a claim or adding extra items to a claim, or being in any way dishonest with the intention of

gaining more than legitimate entitlement.”This definition applies to insurance fraud com-

mitted by fraudulent policyholders. Our paper will focus on the analysis of fraud behavior as

such. Although the types of insurance fraud can be categorized by the nature of insurance

product, in a broader context, the Insurance Council of Australia (ICA) categorizes insurance

fraud as opportunistic, premeditated, or fraudulent nondisclosure, either of which ultimately

leads to illegitimate financial gain for the fraudulent claimant. Present discourse involving

fraud classification on moral grounds may take different positions, which is not pertinent to the

modeling of fraud (Derrig, 2002). Thus, for axiomatic reasons, this paper focuses on the broader

notion of insurance fraud in accordance with ICA, without such further classification.

Traditional insurance fraud detection heavily relied on adept auditing and inspection (Artís

et al., 2002; Dionne et al., 2009; Nian et al., 2016). However, because of technological devel-

opments and large‐scale business operations, the adoption of such conventional methodologies

makes the task of fraud detection impractical (Kemp, 2010). The modus operandi of criminals,

which is ever so dynamic, has made it improbable, if not impossible, to identify fraudulent

claims based on a fixed algorithmic criterion most of the times (Nguyen & Perez, 2020). In his

paper on a holistic approach to fraud management, Wilhelm (2004) discussed this dynamic

behavior of criminals in general to explain why fraud management life cycle should be more

dynamic and evolving. To keep up with the rapid macroenvironmental changes, the insurance

industry had to integrate automation of expeditious statistical models into their internal fraud

detection systems. A study by the Statistical Analysis System found that in the United States,

75% of all insurers had integrated automated systems for fraud detection by 2016.

Sharma and Panigrahi (2013) categorized the key domains in which these models operate as

follows: classification (Phua et al., 2010), clustering (Jain et al., 1999), regression (Fawcett &

Provost, 1997), visualization based on trees, and anomaly‐based (Noble & Cook, 2003). A list of

fraud modeling methodologies is summarized in Table 1. An extraneous observation is the

TABLE 1 Fraud modeling methodologies

Model class Modeling methodologies

Clustering K‐means clustering

Nearest neighbors

Classification Random forest

Naive Bayes

Support vector machines (SVM)

Regression Logistic regression

Polynomial regression

Neural networks

Outlier detection Isolation forest

Gaussian mixture models

Prediction Neural networks

592

|

GOMES ET AL.

possibility that under such a categorization, certain models can share several domains, for

example, tree‐based visualization approaches against classification and regression domains.

Various statistical models have been applied to insurance sectors. See Li et al. (2008) for a

survey on statistical methods for health care fraud detection. Despite varying model categories

and the range of statistical models studied in fraud modeling literature, effectiveness in real‐life

business settings has been the major limitation owing to various reasons, for example, class

imbalance of the data distribution and the inexistence of true outcomes in fraud con-

textualization, which render modeling attempts through conventional statistical frameworks

difficult. In this data mining era with abundant availability of data, the use of machine learning

algorithms is a propitious alternative.

Machine learning algorithms used for fraud detection can be primarily divided into the fol-

lowing categories: supervised (Khatri et al., 2020), unsupervised (Srivastava & Salakhutdinov, 2014),

and semisupervised models (Van Engelen & Hoos, 2020). The significance of such a distinction is

best understood by analyzing the elemental fundamentals of the two initial model classes.

Supervised learning is defined as attempting to infer a function that best maps the given input data

to a given output. A set of training data is submitted as system input during the model training

phase. Each input is labeled with a desired output value, which essentially supervises the model;

therefore, the model knows what the output is when input is fed into it. A review of fraud detection

literature revealed that the most common and pertinent studied class of models are supervised

machine learning models. A recent study of antifraud methodologies used for health care fraud

detection by the Society of Actuaries found that almost all of the methodologies used are supervised

models. Supervised learning paradigms have been well refined over the years through extensive

academic research with significant model developments. Comprehending the analytical framework

of the supervised models is imperative not only to evaluate performance against deep learning

models, but also to understand the classification dynamics of supervised models with the ex-

pectation of better distinguishing model results for greater insights. However, such a discussion is

beyond the scope of this paper.

To develop a supervised model for fraud detection, we must first obtain past information on

the true occurrence of fraud and nonfraud. Based on the accuracy of the output labels, the

model attempts to optimize the parameters to better identify fraud and nonfraud instances. In

contrast, unsupervised learning comprises a class of analytical methods that attempt to infer a

function that best characterizes the representation of the input data in the absence of any

supervision. Unsupervised models based on neural networks with multiple “hidden”layers,

also known as multilayer perceptrons (MLP), are often referred to as unsupervised deep

learning (Goodfellow et al., 2016; LeCun et al., 2015). Hastie et al. (2009) described un-

supervised learning to be more subjective than supervised learning, as there is no simple goal

for analysis, such as predicting a response. Thus, it is able to better determine a feature

representation of the data in the absence of any bias or misguidance precipitated by subjective

output data. An unsupervised learning paradigm plays an important role in fraud detection

within the domain of anomaly detection. Unlike supervised models, we do not require prior

knowledge of the output labels to develop a model (Niu et al., 2019). See Chalapathy and

Chawla (2019) for a review of deep learning methods for anomaly detection.

It is clear from the very definition that supervised modeling is proficient at modeling

relationships in which the output label or variable of interest is naturally generated without the

need for self‐verification. An example is modeling stock prices based on historical book ratios

and stock statistics. The variable of interest in this case, the stock price, is naturally present.

However, when attempting to monitor insurance fraud, the predictive variable of interest is not

GOMES ET AL.

|

593