Diagnosing Model Performance Under Distribution Shift
Authors:
- TiffanyTITiffany
- ) Cai)C) Cai
- Hongseok NamkoongHNHongseok Namkoong
- Steve YadlowskySYSteve Yadlowsky
Preprint
Abstract
Prediction models can perform poorly when deployed to target distributions different from the training distribution. To understand these operational failure modes, we develop a method, called DIstribution Shift DEcomposition (DISDE), to attribute a drop in performance to different types of distribution shifts. Our approach decomposes the performance drop into terms for 1) an increase in harder but frequently seen examples from training, 2) changes in the relationship between features and outcomes, and 3) poor performance on examples infrequent or unseen during training. These terms are defined by fixing a distribution on $X$ while varying the conditional distribution of $Y \mid X$ between training and target, or by fixing the conditional distribution of $Y \mid X$ while varying the distribution on $X$. In order to do this, we define a hypothetical distribution on $X$ consisting of values common in both training and target, over which it is easy to compare $Y \mid X$ and thus predictive performance. We estimate performance on this hypothetical distribution via reweighting methods. Empirically, we show how our method can 1) inform potential modeling improvements across distribution shifts for employment prediction on tabular census data, and 2) help to explain why certain domain adaptation methods fail to improve model performance for satellite image classification.
Keywords
All Related Versions
This publication has 55 references indexed in Scilit:
- External Validation of a Widely Implemented Proprietary Sepsis Prediction Model in Hospitalized PatientsJAMA Internal Medicine, 2021
- Balancing vs modeling approaches to weighting in practiceStatistics in Medicine, 2020
- Causal Inference for Statistics, Social, and Biomedical Sciences2015
- Cross-Validated Targeted Minimum-Loss-Based Estimation2011
- A warm-start approach for large-scale stochastic linear programsMathematical Programming, 2009
- Fertility and female labor supply in Latin America: New causal evidenceLabour Economics, 2007
- Improving predictive inference under covariate shift by weighting the log-likelihood functionJournal of Statistical Planning and Inference, 2000
- Convergence rates and asymptotic normality for series estimatorsJournal of Econometrics, 1997
- Importance measures in global sensitivity analysis of nonlinear modelsReliability Engineering & System Safety, 1996
- Chapter 36 Large sample estimation and hypothesis testing1994