Learning the Distribution Shift in Performative Prediction

Oregon Convention Center 

In many prediction problems, the predictive model affects the sampling distribution. For example, job applicants may doctor their resumes to get them past resume screening systems. Such distribution shifts are especially common in social computing, but methods for learning such distribution shifts from data are scarce. Drawing inspiration from a microeconomic model that characterizes agents in labor markets, we introduce a novel approach for learning the distribution shift. This approach assumes that the agents in the population are rational and adapt to the predictive model by choosing from a finite set of actions. We demonstrate, both empirically and theoretically, that learning the distribution shift and plugging it into an existing method for performative prediction, which necessitates knowledge of the distribution map, is (statistically) more efficient than zeroth-order methods that avoid learning the distribution map.

Strategic classification

Performative prediction

Distribution shift

Cost estimation 

Section on Nonparametric Statistics