How Machine Learning Helps Prevent E-Commerce Fraud
TransUnion Blog, Glen Goldstein | Aug. 22, 2018
Sometimes it takes a good computer to thwart a bad actor
Machine learning, which essentially programs a computer to learn, is already in use all around you. This technology is what picks your recommendations on Netflix and suggests items for you at Amazon. Primitive versions of machine learning have been around for over 60 years, and modern iterations are being used to sort through massive amounts of data to learn how to detect fraud in ways that humans or basic software cannot easily replicate.
Traditional computer algorithms use rules-based systems comprised of a fixed set of formulas that achieve a desired result. A simple example would be a conditional if-then statement like if an e-commerce account uses more than 3 credit cards for more than 10 purchases, then the system should send an alert about potential fraud.
Well-designed examples of these types of algorithms can be very effective at detecting relevant behaviors, but they’re limited by their fixed, initial programming by humans. That can mean new threats or surprising factors that signal the emergence of fraud may not be caught by traditional rules-based detection.
What makes machine learning different from rules-based systems?
Machine learning algorithms are designed to learn from experience and iteratively improve their performance. Essentially, they adapt as the massive number of factors and conditions evolve within a data set, and can essentially keep up with many of the behaviors that humans—and in the case of e-commerce fraud, criminals—throw at them.
For example, let’s say a high degree of fraudulent activity has been recently tied to transactions originating from a geographic area and certain types of credit card accounts, at specific transactional volumes for specific products or services. A fraud detection solution that uses machine learning would iteratively determine these unique combinations warrant a red flag.
By analyzing vast amounts of data and learning the patterns of normal behavior, these programs identify when potentially abnormal activity or data is present — resulting in more accurate detection of fraud. These systems examine data elements that include:
- The device employed for the transaction
- The identity attached to the device
- How the device behaves as it moves through the process
- The identity of the customer
- The customer’s transaction history
- The reputation of the device
- A vast array of other factors and behaviors
This is all driven by an underlying machine learning process that gets more effective as it operates, adapting to fraud in ways that are superior to traditional, rules-based systems.
Since fraudsters are always changing their methods and approaches, businesses need an agile system that becomes more effective with each transaction.
Fraud will never go away, and as we use the Internet for more transactions, retailers need to develop new methods to combat constantly-evolving fraud. Machine learning can help meet these new threats, protecting your customers, business and bottom line.