‘Large-Scale Machine Learning’ Drives Sift Science’s Antifraud Efforts

March 21, 2013

Sift Science, a Y Combinator-backed startup, launched this week with what it believes is a novel approach to fighting e-commerce fraud. Unlike many antifraud solutions that use a set of fixed rules to identify fraud, the San Francisco-based company uses “large-scale machine learning” to identify shifting behavioral patterns that suggest fraud.

A bit of Javascript code tracks user behavior at a Website and passes the data to Sift Science, which generates a fraud score to assess the risk of a transaction based on past events. Each e-commerce Website that is part of the Sift Science network contributes data and the system constantly learns more patterns that lead to fraud, making the solution increasingly effective, said Brandon Ballinger, the company’s co-founder and CTO.

“What makes large-scale learning unique is the detail of patterns learned,” Ballinger wrote on the Sift Science blog this week. “Like peering through a microscope, when you up the resolution, you can spot surprising details the naked eye would never notice. For example, a user who signs up and waits an hour before making a purchase is 7x more likely to generate a chargeback than a user who purchases immediately after signup.

“Our system has pinpointed particular page navigation sequences, IP ranges, email address patterns, graph connectivity structures, browser configurations, and even types of text entered that predict fraudulent activity. And it’s learning more patterns each day,” he added. “Fraudsters don’t play by a fixed set of rules. So why should you?”

The company employs several ex-Google employees and Google vet Ballinger has characterized Sift Science’s technology as similar to what Google uses to recognize speech, rank search results and show ads. The fraud-fighting startup received a $4 million Series A funding round in November of 2012.