Detecting Fraud Using Descriptive, Predictive, and Social Network Techniques

A typical organization loses 5% of its revenues to fraud each year, and, in the U.S. alone the total cost of (non-health) insurance fraud is estimated at more than $40 billion per year, according to Bart Baesens, one of the authors of a new book on data science techniques for fraud detection. While opportunities for fraud are continuously evolving, fraud detection always has the biggest impact if it is done quickly. Fortunately, says Baesens, big data analytics can help to make that possible.

What is different about companies’ ability to effectively deal with fraud now versus the past? Is it the nature of the crime or the tools for detecting it?

Baesens: Actually, I think it’s both.  In terms of tools, we can now fight fraud using state-of-the-art analytical techniques implemented in a big data environment.  One of the most powerful information sources for detecting fraud concerns social networks.  In an insurance setting, you can think of a network of nodes such as claimant, policyholder, car, car repair shop, etc.  We can now analyze the complex relations between all these nodes to unveil sophisticated collusion patterns and fraud rings.  In our own research, we applied social networks to both credit card and tax avoidance fraud and found some amazing lifts compared to traditional fraud detection models.  Unfortunately, as the techniques become more sophisticated, so do the fraudsters so it’s important to always stay one step ahead and continuously back-test and perfect your analytical fraud models.

What prompted you to write this book?

Baesens: A typical organization loses 5% of its revenues to fraud each year. For instance, the total cost of insurance fraud (non-health insurance) in the U.S. is estimated to be more than $40 billion per year. So, there is a significant cost-saving potential for fighting fraud. Providing the tools to do this is the aim of this book. 

More specifically, we strongly believe that using state of the art analytics is the most efficient way to detect fraud as soon as possible and as accurate as possible, and since (big) data is becoming available in abundance and at a low cost in many business settings, fraud analytics are within reach of any organization – at least with the right skills available.  This book teaches them. 

Who is it targeted at?

Baesens: The book describes the data necessary to detect fraud, and then takes the reader from the basics of fraud detection data analytics, through advanced pattern recognition methodology, to cutting edge social network analysis and fraud ring detection.  It is targeted at anyone working in fraud detection and prevention across a wide range of industries such as financial services, insurance, Telco, pharmaceutical, and government. Also fellow researchers such as colleague professors or aspiring Ph.D. students might find it useful. 

You make the point that catching fraud early is key to limiting the damage. How far along is fraud typically noticed and halted?

Baesens: Fraud detection is more valuable the sooner it is done, because further losses are prevented, potential recoveries are higher, and security issues can be addressed more rapidly, as such avoiding cascading damage to the organization. Detecting fraud in an early stage however is harder than detecting it in an evolved stage, and requires specific techniques which we also cover. 

The speed with which fraud must be detected depends upon the business setting you are working in.  For example, in a credit card business setting, a fraud decision must be made in less than 5 seconds after the initiation of the transaction.  Also in online environments such as online payments and click fraud time is of utmost importance.  This time pressure puts huge challenges on the analytical models to be implemented.  In other settings such as insurance fraud and tax evasion, time pressure is less of an issue and more resources can be spent to thoroughly investigate the claims and transactions. 

Overall, is there any statistical evidence that fraud is increasing or has changed in its frequency?

Baesens: I believe it’s hard to say, since both the prevalence and intensity needs to be considered which are hard to compare in a fast-changing environment where new fraud opportunities arise on a continuous basis.  As an example, consider the Internet of Things referring to the network of interconnected things such as electronics devices, sensors, software, and IT infrastructure. In this setting, fraudsters might force access to web-configurable devices like ATMs and set up fraudulent transactions.  Another example is device hacking whereby fraudsters change operational parameters of connected devices such as manipulating smart meters to make them under register actual usage.  To answer your question, I believe fraud has increased, but the real challenge is that it continuously changes and adapts to its environment. 

What do companies need to do to more effectively thwart fraud? 

Baesens: I would say invest in people, data, and analytics, in that order. 

Staff needs to be educated and trained to understand how state of the art technology can be used to detect and prevent fraud.  Since fraud is a key treat to a firm’s revenues and consequently its existence, it is important that senior management and the board of directors are also actively involved in creating this awareness and engagement.  More specifically, they should demonstrate active involvement on an ongoing basis, assign clear responsibilities, and put into place organizational structures, procedures and policies that will allow the proper and sound management of fraud. 

Next, corporation-wide data platforms need to be set up to provide a comprehensive 360-degree view on customer behavior.  Remember, data is the key ingredient to any fraud analytical model and it is of crucial importance that this data is readily available, at low cost, and with the required minimum quality. 

Finally, once the data architecture has been put in place, analytical models need to be developed to understand the complex patterns in it.  These detected patterns can then be deployed into sophisticated fraud prevention strategies to protect the firm against the threats it is continuously exposed to. 

“Fraud Analytics Using Descriptive, Predictive, and Social Network Techniques: A Guide to Data Science for Fraud Detection” is written by Bart Baesens, Veronique Van Vlasselaer, and Wouter Verbeke.

Image courtesy of Shutterstock

Related Articles

With increased demand for mass customization and personalization, the emergence of Web 2.0, and one-to-one marketing, and the need for better risk management and timely fraud detection, the pressure is on for organizations to improve their ability to extract, understand, and exploit analytical patterns of customer behavior and strategic intelligence.

Posted November 12, 2015


Subscribe to Big Data Quarterly E-Edition