Rules vs. Machine Learning: Why You Need Both to Win
By Kevin Lee /
12 Jul 2018
Rules-based fraud prevention systems have a number of drawbacks when compared to the sophistication of machine learning, but is a machine learning-based system the be-all and end-all of fraud prevention? Can rules actually prove to be beneficial while making the transition to machine learning?
We explored these and other questions in our webinar, Rules, Machine Learning, and the Best Way to Fight Fraud. The short answer is that while machine learning stands head and shoulders above the capabilities of rules, it would be a mistake to consider machine learning a silver bullet to all of your company’s fraud woes.
Let’s start from the top – how does machine learning improve upon rules?
- Rules can be reverse engineered.
- Ex: Using a binary rules system, an ecommerce company sets a rule that any purchase $200 and above will be declined. Fraudsters might make a purchase for $199, which won’t get flagged, causing the company to reset the purchase limit that will be declined. This can become a neverending game of cat and mouse.
- As rules become stricter, you risk negatively impacting good customers.
- Integration of rules is time-intensive and costly.
- Rules require more upkeep, and new rules will need to be made as time goes on, which requires research and data analysis.
- Performance issues are common when sudden changes are introduced into the system.
- Ex: A big holiday sale happens and your site is hit with an influx of good orders, but a rules-based system can’t adapt quickly to the new behaviors it’s seeing. This can result in declined orders by legitimate customers and a lot of frustrated consumers.
- Rules-based systems open up greater opportunities for human error.
- These systems require a lot of human input, and when humans get overloaded with data they tend to make mistakes, especially if they start to get bored on the job.
Machine learning is a solution to these problems. An automated system doesn’t need to rely on a human to be paying attention to every small detail that comes across their desk. As machine learning systems collect more data and become smarter, the number of false positives decreases. The time invested into upkeeping a machine learning system is a fraction of what’s required for maintenance of rules.
Yet, despite all of its benefits, machine learning is not the silver bullet of fraud prevention. While it solves a lot of the issues that rules-based systems are confronted with, rules and machine learning can actually support one another.
Can rules still be beneficial to a machine learning system?
If your company is transitioning from rules to machine learning, it’s important to note that machine learning models take time to acclimate and learn from data. A machine learning model is only as good as the data it has absorbed, and it can take months for the system to be ready to replace rules.
That being said, it’s recommended that rules run alongside machine learning during that time. When integrating Sift Science, companies can incorporate Sift Scores into existing rules and begin using the console, and then phase out rules over time as the machine learning model proves its accuracy.
It’s also worth considering whether operating rules and machine learning systems in tandem would be more beneficial to your company than replacing rules entirely. Machine learning isn’t meant to substitute for humans, but instead to augment what we’re capable of. While machine learning’s strength is in the amount of data it can analyze and track in real time, a human’s strength is in providing context and intuition and analyzing edge-case scenarios.
For example, let’s say I make a purchase using the email email@example.com and a machine learning model flags it as suspicious because it doesn’t contain “Kevin Lee” or isn’t at all remotely similar in vowel/consonant construction. But a human reviews the flag and notices that my IP or shipping address is based in Chicago. The White Sox are a Chicago team, Michael Jordan played for the Chicago Bulls, and his number was 23. Machine learning wouldn’t be able to pick up on those details, which can make a human’s intervention the difference between a legit transaction being approved and a false positive.
Both rules and machine learning have their respective pros and cons, and while machine learning offers some amazing solutions to fighting fraud, it would be unwise to think of it as a panacea.
If you want a deeper dive into rules versus machine learning, be sure to watch our free webinar!
Kevin Lee is Vice President of Digital Trust & Safety at Sift. Building high-performing teams and systems to combat malicious behavior are what drive him. Prior to Sift, Kevin worked as a manager at Facebook, Square, and Google in various risk, spam, and trust and safety roles.