Performative Prediction with Selective Labels
The work introduces performative prediction, where models are optimized for a training distribution but deployed under a different one. It addresses the 'Selective Performative Gap,' where retraining only on observed data can lead to unstable solutions. The authors propose Robust Repeated Risk Minimization (Robust RRM), which uses worst-case loss estimates based on confidence intervals to ensure convergence toward a stable solution, even with incomplete label information.