Performance Prediction Under Dataset Shift

TL;DR

This paper evaluates how well performance prediction models generalize to unseen domain shifts, finding that error predictors outperform shift detection metrics and proposing an uncertainty estimation method for reliable predictions.

Contribution

It demonstrates the limitations of shift detection metrics and introduces error predictors with uncertainty estimation for better performance prediction under dataset shift.

Findings

Error predictors outperform shift detection metrics on unseen domains.

Uncertainty estimation improves the reliability of performance predictions.

Models trained on synthetic shifts generalize better to real domain changes.

Abstract

ML models deployed in production often have to face unknown domain changes, fundamentally different from their training settings. Performance prediction models carry out the crucial task of measuring the impact of these changes on model performance. We study the generalization capabilities of various performance prediction models to new domains by learning on generated synthetic perturbations. Empirical validation on a benchmark of ten tabular datasets shows that models based upon state-of-the-art shift detection metrics are not expressive enough to generalize to unseen domains, while Error Predictors bring a consistent improvement in performance prediction under shift. We additionally propose a natural and effortless uncertainty estimation of the predicted accuracy that ensures reliable use of performance predictors. Our implementation is available at https: //github.com/dataiku-research/performance_prediction_under_shift.