Using Machine Learning in Rules-Based Fraud Prevention Systems

Machine learning models, which essentially give computer systems the ability to “learn” to perform a specific task without being explicitly programmed to do so, are getting smarter. As such, we’re starting to find them everywhere: in Netflix’s prediction algorithms, art history studies, and even in medical diagnostic software.

And, of course, in fraud prevention. Machine learning is the latest “new thing” in eCommerce, and a lot of merchants are curious about whether it will replace a rules-based system entirely. Instead, I think we should be asking how to utilize machine learning within a rules-based system.

How is machine learning different than a rules-based system? In a nutshell, a rules-based system must be manually created by humans, whereas a machine learning model comes to its own conclusions based on patterns it finds in data. Machine learning and rules-based systems each have their own pros/cons, which can complement each other. In a tutorial presentation on machine learning, Laura Chiticariu of IBM lists the benefits and limitations of each:

You’ll notice many of the limitations of one system are balanced out by the advantages of the other. For example, where a rule-based system is hands-on and requires constant manual labor, a machine learning system adapts automatically to reduce the amount of human time and effort required. On the other hand, whereas a machine learning system can be opaque and require machine learning expertise to maintain, a rules-based system is much easier to comprehend, tweak, and debug.

We encourage merchants to use an approach that combines both machine learning and a rules-based system. In fact, we recently found a combination strategy to be effective in reducing fraudulent orders of our consumer product, Whitepages Premium.

Adding Value with Machine Learning

The most valuable use of using machine learning is in combining multiple scores to make strong accept and/or reject rules. In the lending world, they’ve been doing this for years, using scores from all three of the credit bureaus (Equifax, Experian, and TransUnion), to evaluate a consumer’s credit risk. For example, if all the scores are above 800, the lender feels very confident on lending to that borrower.

You can use eCommerce machine learning scores in a similar way. A common example I see across merchants is utilizing two to three machine learning scores to create accept or reject rules in their decisioning platform. Each score has its own network and data sources that allows merchants to be more confident in their decisions by corroborating the data.

For example, Whitepages Pro offers the Confidence Score as part of our Identity Check API, which uses machine learning to score transactions on a scale of 0-500. We have found transactions with scores 450 and over highly correlate with fraud, which means a merchant can catch fraud they may have otherwise missed with just their rule set. For instance, one top office supply retailer found $75k in missed fraud in a month when implementing the Confidence Score.

Machine learning scores from different sources can be combined to create even stronger rules. Merchants using Whitepages Pro in conjunction with CyberSource, for example, can write rules around the Confidence Score and CyberSource’s Advanced Fraud Screen (AFS) score such as:

  • If the AFS Score is below 40 AND WPP Confidence Score is below 150 —> accept
  • If the AFS Score is greater than 95 AND WPP Confidence Score is greater than 475 —> reject or review

As well as finding fraud, merchants also see lift with the Identity Check Confidence Score in reducing manual review by creating modifications to review rules. For example, if you see a billing/shipping mismatch rule has a 90%+ acceptance rate from the manual review team, you can add the condition “AND WPP Confidence Score is greater than or equal to 150.” In this case, it will remove all transactions with a score less than 150 from that specific review rule.

As with any fraud rule, those based around machine learning scores will still need to be monitored and adjusted as needed. Fraud is a balancing act and rules should be constantly checked to make sure they’re meeting your company’s KPIs. For example, if you have bandwidth on your manual review team, you may create more review rules to pass transactions to your team and tighten up your fraud prevention program even further. Machine learning models are simply another powerful tool to help make these rules even more effective and efficient.

To evaluate how the Confidence Score could provide lift to your current fraud KPIs, get a free data evaluation from one of our experts.

Thanks for reading! You might be interested in these posts, too: