Skip to content Skip to navigation

Research & Ideas

Search this site

​Claudia Perlich: The secret life of predictive models

​At the second annual Women in Data Science conference, an expert in the field says data scientists must be careful what they ask for — and wary of what they find.

Predictive model in question mark

Predictive models have lead to tremendous advances, but they’re not without their pitfalls. | Illustration by Stefani Billings

 

“I fell in love with predictive modeling more than 20 years ago,” said Claudia Perlich, chief scientist of Dstillery, during a presentation at the second annual Women in Data Science conference. “It was a way for the eternal introvert in me to understand the world without having to talk to people.”

Despite her longstanding affection for predictive models and the tremendous advances they’ve made, Perlich provided a warning that they can lead you astray if you’re not careful. In particular, machine learning is susceptible to what she calls the “certainty vortex” of chasing the wrong signals and missing what you’re really looking for. Whether it’s online advertising that targets bots rather than human browsers or policing and hiring algorithms that lead to unintentional discrimination, Perlich said that “there is some secret life to the models we build, and it is our duty to get to the bottom of it.” She cautioned against letting predictive models make important decisions alone. “I see this as a collaboration, as a second opinion,” she said. “I’m always scared when a model gets too good. There’s usually something wrong.”

Get Updates from Stanford Engineering