Detecting and Correcting for Label Shift with Black Box Predictors

By Zachary Chase Lipton, Yu-Xiang Wang, Alex Smola. 2018

671 citations recorded by the source corpus.

Faced with distribution shift between training and test set, we wish to detect and quantify the shift, and to correct our classifiers without test set labels. Motivated by medical diagnosis, where diseases (targets), cause symptoms (observations), we focus on label shift, where the label marginal $p(y)$ changes but the conditional $p(x|y)$ does not. We propose Black Box Shift Estimation (BBSE) to estimate the test distribution $p(y)$. BBSE exploits arbitrary black box predictors to reduce dimensionality prior to shift correction. While better predictors give tighter estimates, BBSE works even when predictors are biased, inaccurate, or uncalibrated, so long as their confusion matrices are invertible. We prove BBSE's consistency, bound its error, and introduce a statistical test that uses BBSE to detect shift. We also leverage BBSE to correct classifiers. Experiments demonstrate accurate estimates and improved prediction, even on high-dimensional datasets of natural images

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Safety, Robustness & Interpretability

Citations in the corpus

Learning Robust Global Representations by Penalizing Local Predictive Power (2019, background): “…(when p ( y | x ) does not change) (Gretton et al., 2009) and under label shift (when p ( x | y ) doesn’t change) (Storkey, 2009; Zhang et al., 2013; Lipton et al., 2018), and various other assumptions (e.g. bounded divergences between source and target distributions) (Mansour et al., 2009; Hu et…”
Evaluating Models’ Local Decision Boundaries via Contrast Sets (2020, background): “…strikingly evident when models achieve high test accuracy but fail on simple input perturbations (Jia and Liang, 2017; Feng et al., 2018; Ribeiro et al., 2018a), challenge examples (Naik et al., 2018), and covariate and label shifts (Ben-David et al., 2010; Shimodaira, 2000; Lipton et al., 2018).”
AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (2019, background): “In practice, data distributions evolve (Lipton et al., 2018), models encounter new scenarios (Hendrycks & Gimpel, 2017), and data curation procedures may capture only a narrow slice of the underlying data distribution (Torralba & Efros, 2011).”
Toward Causal Representation Learning (2021, background): “It also led to a range of follow-up work [32], [78], [97], [114], [115], [152], [153], [156], [167], [195], [204], [243], [263], [267], [277], [278], [281], complementing the studies of Bareinboim and Pearl [14], [185], and it inspired a thread of work in the statistics community exploiting invarian”
Causality for Machine Learning (2019, background): “The line of work provided rather fruitful (Zhang et al., 2013; Weichwald et al., 2014; Zhang et al., 2015; Gong et al., 2016; Huang et al., 2017; Blöbaum et al., 2016; Guo et al., 2018; Wang et al., 2019; Subbaswamy et al., 2018; Rabanser et al., 2018; Li et al., 2018a; Lipton et al., 2018; Li et al”
Failing Loudly: An Empirical Study of Methods for Detecting Dataset Shift (2018, methodology): “(a) Adversarial (adv): We turn a fraction δ of samples into adversarial samples via FGSM [13]; (b) Knock-out (ko): We remove a fraction δ of samples from class 0, creating class imbalance [29]; (c) Gaussian noise (gn): We corrupt covariates of a fraction δ of test set samples by Gaussian noise with ”
Learning Robust Representations by Projecting Superficial Statistics Out (2018, methodology): “…2000; Gretton et al., 2009), and label shift (also known as target shift or prior probability shift ) (Storkey, 2009; Sch¨olkopf et al., 2012; Zhang et al., 2013; Lipton et al., 2018), principled methods exist for estimating the shifts and retraining under the importance-weighted ERM framework.”
Domain Adaptation with Asymmetrically-Relaxed Distribution Alignment (2019, background): “Some consider the reverse setting label shift (Saerens et al., 2002; Zhang et al., 2013; Lipton et al., 2018), where although the label distribution shifts ( p S ( y ) (cid:54) = p T ( y )), the class-conditional input distribution is invariant ( p S ( x | y ) = p T ( x | y )).”
Multiaccuracy: Black-Box Post-Processing for Fairness in Classification (2018, background): “In particular, multiaccuracy post-processing may be applicable in domain adaptation settings, particularly under label distribution shift as studied recently in [22], but when the learner gets a small number of labeled samples from the new distribution.”
Evaluating NLP Models via Contrast Sets (2020, methodology): “This is strikingly evident when models achieve high test accuracy but fail on simple input perturbations (Jia and Liang, 2017; Feng et al., 2018), challenge examples (Naik et al., 2018), and covariate and label shifts (Ben-David et al., 2010; Shimodaira, 2000; Lipton et al., 2018). ar X iv :2 00 4.”
WILDS: A Benchmark of in-the-Wild Distribution Shifts (2020, neutral): “…e.g., learning to map data points from source to target domains (Hoffman et al., 2018), or estimating the test label distribution from unlabeled test data (Saerens et al., 2002; Zhang et al., 2013; Lipton et al., 2018; Azizzadenesheli et al., 2019; Alexandari et al., 2020; Garg et al., 2020).”
ReAct: Out-of-distribution Detection With Rectified Activations (2021, background): “Meanwhile, some works considered label distribution shift [1, 2, 32, 42, 47, 53], where the label space is common between ID and OOD but the marginal label distribution changes, as well as covariate shift in the input space [14, 39], where inputs can be corruption-shifted or domainshifted [17, 48].”
Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss (2019, methodology): “We are addressing a largely different question: can we do better than re-weighting or re-sampling when the label shift is known? In fact, our algorithms can be used to replace the re-weighting steps of some of the recent interesting work on detecting and correcting label shift [36, 3].”
A Survey on Negative Transfer (2020, background): “PS(Y |X) = PT (Y |X) PS(X) , PT (X) 4) Covariate Shift [28]: The feature conditional distributions are the same, but the marginal distributions of the features are different, i.”
LTF: A Label Transformation Framework for Correcting Label Shift (2020, methodology): “In the discrete Y scenario, Lipton et al. (2018) proposed a method which estimates the importance weight ( P TY / P SY ) by matching the output of trained classiﬁer on the training set (confusion matrix), and then Azizzadenesheli et al. (2019) turned this problem as a linear programming problem and…”
A Review of Domain Adaptation without Target Labels (2019, methodology): “importance weights without requiring target labels [8], [81], [113].”
Fairness without Demographics through Adversarially Reweighted Learning (2020, background): “Domain adaptation work requires a model to be robust and generalizable across different domains, under either covariate shift [53, 48] or label shift [40].”
Towards Out-Of-Distribution Generalization: A Survey (2021, background): “There are multiple taxonomies for the OOD generalization problem [154, 223, 246, 27].”