From Sift Science
When it comes to user fraud, marketers should push for machine learning systems to replace rule-based ones.
That’s the viewpoint of Jason Tan, CEO/co-founder of San Francisco-based Sift Science. Of course, it would be shocking if he didn’t feel that way, since his company — started by several former Google engineers — is one of several startups that are now offering machine learning-based user fraud detection.
He recently made the case to me that machine learning should be on every marketer’s wish list.
Because, he said, rules-based systems throw up hoops to users who are trying to buy something or trying to register. These hoops include captchas, those squiggly images of letters and numbers that humans should be able to decipher and which software agents — bots — cannot.
Hoops can also mean security questions or reviews by human analysts. Fraudsters can include humans wielding stolen credit card numbers.
As every marketer knows, the more hoops you make users jump through to buy something or sign up, the less chance they will.
“Most of the hoops are unnecessary,” Tan said, adding that they’re “mostly there because of existing anti-fraud systems that are rules-based.” And the hoops or delays can be an even bigger issue on mobile, he noted, where e-commerce and registration completion rates are lower than on desktop.
Except for big players like Google, Amazon, PayPal and Facebook, Tan noted, almost all of the current anti-fraud solutions out there are rules-based.
“What [Sift Science] has built is what Google has had for ten years,” he said, adding that machine-learning systems will eventually take over fraud detection from the rules watchers.
“The Way A Human Learns”
Another machine-learning anti-fraud firm, Vancouver-based NuData Security, notes on its website that the hundreds of rules can include the direction to “flag any purchase over $40,000 [or] flag out-of-state shipping.” Flagged users are then usually scanned manually by anti-fraud analysts.
Rules-based systems are committed to an escalating war with fraudsters, who find the weaknesses and exploit them. Then the systems add more rules.
By contrast, machine learning looks at fraudulent cases and, NuData says, the algorithms automate “the investigation of more data than is possible for a human to screen, correlating hundreds of data points like typing speeds, scrolling speeds, preferred times of day to visit, top cities, countries, devices, credit card numbers, credit card types” and so on.
Sift Science adds such factors as email addresses, which it says it analyzes “six ways to Sunday.” Tan notes that the fraudster often doesn’t see the patterns, and machine learning will adapt if the patterns change.
A machine learning system, Tan noted, “is able to learn from data the way a human learns from experience.” And it does it in real time.
He claims that machine learning-based anti-fraud systems are “five to ten times better” than rules-based ones, and they don’t set up all those hoops. With rules-based systems, he said, businesses have to choose one of the following, because the “other two will suffer”:
A low fraud rate
Effective review operation
A frictionless customer experience
“The big advantage to machine learning,” he said, “is that you don’t have to choose one.”
Sift Science says that one of its clients, internet retailer JackThreads, found that 92 percent of the users it detected with machine learning as fraud actually were, while the industry average for rules-based systems is 25 percent.
In general, Tan says, rules-based systems flag anywhere from ten to 50 percent of users for human review, while machine learning systems single out less than one or two percent.
If you want to get a digital gift card, a review within a rules-based system might mean you have to wait a day or two, he said. But with real-time machine learning detection, you can get your card in seconds.
Comments