The pros and cons of using machine learning to prevent fraud

It’s that time of the month again. Your bank statement is ready, and you’re about to begin the sobering task of reviewing your expenses. Skimming through the transactions, your eyes stop abruptly at a peculiar entry: a US$700 charge on Amazon. You try to recall making this transaction, but the truth is you’ve become a victim of fraud.

Long ago, racketeers had to physically steal credit cards to make fraudulent purchases. As shopping evolved, so did the scammers. With the expertise to harvest personal information from the Internet, they can now plunder and sell thousands of identities easily with just a few clicks.

Personal data that is needed for ecommerce fraud, for instance, can be hacked from unencrypted non-ecommerce websites. Each piece of information is matched and put together to form a complete identity, which can then be used to commit the crime.

“It’s really important for consumers to have a good understanding of the websites they go to or the apps they use on their devices, and where any information they put into these apps and websites could wind up going to,” shares Scott Boding, vice president of risk solutions product management at payment management firm CyberSource.

No matter how cautious you are, though, there are always more sophisticated perpetrators who can outwit you. This is where machine learning can swoop in to save the day.

What is machine learning?

Machine learning, a subdivision of AI, enables machines to independently process and learn from past information to predict future patterns. It is already being used in the development of autonomous cars and in medical studies. In such scenarios where the activities are relatively consistent and predictable, it makes sense to use a static set of data for machine learning applications.

In the ever-evolving territory of ecommerce, however, relying on static data won’t suffice.

“Think about playing a board game with an opponent, where not only is your opponent learning and getting better, but the rules of the game are also changing over time,” says Boding. “Because it’s not a static problem, we have to continually retrain these machine learning systems and keep adding new capabilities so they can continue to expand and get more sophisticated.”

Here are a few advantages and points to consider if a company is exploring machine learning systems for fraud prevention.

Photo credit: gajus / 123RF

The pros and cons of fighting fraud with machine learning

Pro: Scale and speed

It’s impossible for human beings to keep up with the exponential rate at which transactions are made in the ecommerce space. Even if the fastest analyst is able to go through 1,000 transactions a day, they may still fall victim to human error and bias. A machine learning system, on the other hand, can handle billions of transactions and respond at lightning speed with absolute accuracy.

“We’re talking about a matter of 200 or 300 milliseconds at the very most, even when it’s doing a very in-depth historical analysis of transactions,” says Boding.

The ability to process information almost instantly also allows the machine learning system to render an analysis in the middle of a transaction: a critical function for preventing fraud on the spot.

Con: Cost

To build a machine learning system in-house, companies will have to hire not just one, but a team of data scientists to design and continually update the system. As there are massive amounts of data involved, businesses also need to invest in data storage and management, which will eat up more network space as well.

If the machine learning model is poorly made, racketeers will be able to bypass it, leading to an onslaught of fraudulent transactions and subsequent chargebacks.

Pro: Reduced manual labor

Having a machine learning system gets rid of the need for manual labor. The work of 100 employees can be done in the same amount of time by a single piece of software.

Rather than spending inordinate amounts of time authenticating each transaction, teams can focus on more valuable high-level assignments. Companies save a ton of effort on otherwise menial tasks, and make workplace processes more efficient.

Con: High levels of mandatory technical expertise

It is crucial to build a proper machine learning model that doesn’t malfunction. To do that, firms need trained experts who are skilled at building such systems, with an in-depth understanding of the unique field of payment fraud.

“When I hire a new data scientist, it can take from six months to a year for them to have a good handle on the ecommerce fraud space,” Boding shares.

A subpar data scientist might make fatal mistakes and overlook critical details, so it may make more sense to leave the heavy lifting to professional service providers.

Photo credit: Andrii Starunskyi / 123RF

Pro: Unbiased analysis

Being human comes with a slew of limitations, including inherent biases. Informed by previous experiences, a person’s bias can cause them to subconsciously analyze transactions inaccurately and miss out on what’s really going on.

“I’ve done this experiment a number of times, where I’ll put a transaction in front of someone and say, ‘It’s your job to review this. Tell me what you think.’ They’ll often say it’s fraud just because I put it in front of them,” Boding reveals.

Unlike human beings, machine learning systems make no assumptions, which means they’re more capable of conducting accurate analyses.

Con: Difficult collection of good data

To build a viable machine learning model with long-term functionality, it requires a great deal of good data – which isn’t easy to obtain.

For datasets to be statistically relevant, there must be a large enough number of transactions, each of which should contain enough data to work with. This includes information about the outcomes of those transactions.

“If a merchant is doing a million transactions a day but doesn’t get any chargeback information, they’re not going to be able to build a model out of that because they won’t know what’s good and what’s bad,” says Boding.

To find out more about how machine learning helps in ecommerce fraud protection, read this whitepaper by CyberSource, an ecommerce payment management firm that provides a full suite of solutions for ecommerce merchants to combat fraud.