Calibration for the (Computationally-Identifiable) Masses

By Úrsula Hébert-Johnson, Michael P. Kim, Omer Reingold, Guy N. Rothblum. 2017

92 citations recorded by the source corpus.

As algorithms increasingly inform and influence decisions made about individuals, it becomes increasingly important to address concerns that these algorithms might be discriminatory. The output of an algorithm can be discriminatory for many reasons, most notably: (1) the data used to train the algorithm might be biased (in various ways) to favor certain populations over others; (2) the analysis of this training data might inadvertently or maliciously introduce biases that are not borne out in the data. This work focuses on the latter concern. We develop and study multicalbration -- a new measure of algorithmic fairness that aims to mitigate concerns about discrimination that is introduced in the process of learning a predictor from data. Multicalibration guarantees accurate (calibrated) predictions for every subpopulation that can be identified within a specified class of computations. We think of the class as being quite rich; in particular, it can contain many overlapping subgroups of a protected group. We show that in many settings this strong notion of protection from discrimination is both attainable and aligned with the goal of obtaining accurate predictions. Along the way, we present new algorithms for learning a multicalibrated predictor, study the computational complexity of this task, and draw new connections to computational learning models such as agnostic learning.

Subfields: Computational Theory and Mathematics, Statistics and Probability, Optimization & Training

How researchers see this paper

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

In Computational Theory and Mathematics

In Statistics and Probability

In Optimization & Training

Citations in the corpus

Prediction-Based Decisions and Fairness: A Catalogue of Choices, Assumptions, and Definitions (2018, methodology): “apply approximate calibration to groups defined by any of the input variables [55].”
Fairness Without Demographics in Repeated Loss Minimization (2018, background): “A number of authors have also studied individual notions of fairness, either through a fixed similarity function (Dwork et al., 2012) or subgroups of a set of protected labels (Kearns et al., 2018; Hébert-Johnson et al., 2017).”
Preventing Fairness Gerrymandering: Auditing and Learning for Subgroup Fairness (2017, methodology): “Independent of our work, Hébert-Johnson et al. [2017] also consider a related and complementary notion of fairness that they call “multicalibration”. In settings in which one wishes to train a realvalued predictor, multicalibration can be considered the “calibration” analogue for the definitions of ”
A comparative study of fairness-enhancing interventions in machine learning (2018, background): “While there are still few fairness-aware algorithms that can explicitly handle multiple sensitive attributes ([14, 22]), all algorithms discussed can handle them if preprocessed as described earlier so that they are combined into a single sensitive attribute (e.g., race-sex).”
Multiaccuracy: Black-Box Post-Processing for Fairness in Classification (2018, methodology): “This connection between subpopulation fairness and learning is also made in [13, 16, 19], albeit for different tasks.”
Fairness in Machine Learning: A Survey (2020, background): “Similarly, and especially for decoupled classifiers, is the challenge of selecting groups: Some can be rarer than others [283] and as such, a balance is needed to ensure groups are as atomic as possible but robust against gerrymandering [140, 170].”
Towards causal benchmarking of bias in face analysis algorithms (2020, background): “The machine learning community is active in analyzing biases of learning models, and how one may train models where bias is mitigated [3, 14, 18, 31, 33, 41, 46, 51, 68], usually by ensuring that performance is equal across certain subgroups of a dataset.”
Decoupled Classifiers for Group-Fair and Efficient Machine Learning (2017, background): “Hebert-Johnson et al. (2017) extend calibration to multi-calibration which requires the classifier to be well calibrated on a collection of sets of individuals, eg, all those described by circuits of a given size.”
Putting Fairness Principles into Practice: Challenges, Metrics, and Improvements (2019, background): “Further, because the number of examples with demographic information is relatively small, a more expansive intersectional evaluation is not feasible [22, 25].”
Online Learning with an Unknown Fairness Metric (2018, background): “Starting with Joseph et al. [2016a], several papers have studied notions of fairness in classic and contextual bandit problems.”
Why Is My Classifier Discriminatory? (2018, background): “Additionally, extending our analysis to intersectional fairness (Buolamwini & Gebru, 2018; Hébert-Johnson et al., 2017), e.g. looking at both gender and race or all subdivisions, would provide more nuanced grappling with unfairness.”
Fairness Under Composition (2018, background): “In particular, we show that the intersectional subgroup concerns (which motivate [13, 9]) are exacerbated by composition.”
Probably Approximately Metric-Fair Learning (2018, result): “In this sense, it is similar to recent works that protect large collections of sufficiently-large groups (Hébert-Johnson et al., 2017; Kearns et al., 2017; Kim et al., 2018).”
Fairness in Machine Learning (2020, methodology): “Some methods transform the constrained optimization problem via the method of Lagrange multipliers [3, 15, 34, 37, 38, 59, 97, 135, 183, 185] or add penalties to the objective [5, 14, 22, 44, 46, 56, 58, 62, 73, 74, 75, 79, 82, 87, 88, 89, 93, 94, 95, 98, 104, 115, 119, 123, 133, 134, 135, 138, 139,”
On the apparent conflict between individual and group fairness (2019, background): “subgroup doesn’t scale well to the large number of possible combinations, and may lead to overfitting [25, 34].”