Multi-Accuracy Boosting
Description
Multi-accuracy boosting is a post-processing method that corrects a trained classifier's systematic errors on subgroups of people, without retraining it. It needs only the classifier's predictions and a small labelled validation set. An auditor, usually a simple model such as a ridge regression or a shallow decision tree, learns to predict where the classifier's predictions are consistently too high or too low, which finds overlapping subgroups without anyone naming them in advance. Each round fits a small correction for the worst subgroup, as in boosting, and the loop stops when the auditor can find no subgroup whose average error exceeds a chosen tolerance. The result is multi-accuracy, meaning unbiased predictions on every subgroup the auditor can identify, which is not the same as equal accuracy across groups, and the audit itself reports where the original model failed.
Example Use Cases
Fairness
A hospital buys a black-box diagnostic model it cannot retrain. Its data team holds a small labelled sample of local patients and runs the audit, which finds that predictions are systematically too low for older women. The team post-processes the model until no subgroup the auditor can find exceeds the agreed error tolerance, then checks that groups the model already served well did not get worse.
A lender uses a vendor's credit-risk model, and its data holds no sensitive attributes. A risk analyst audits the model on a labelled sample of recent applications with a decision-tree auditor, which finds subgroups defined by ordinary features where the errors concentrate. The correction is applied until those subgroups' average error falls within tolerance.
Limitations
- The guarantee covers only subgroups the auditor can identify from the features. If a group's errors follow a pattern too complex for the auditor, or the features carry little trace of group membership, the audit will pass the model while that group is still poorly served.
- Multi-accuracy bounds the average prediction error within each identified subgroup. It does not equalise error rates, so large gaps can remain. In the original paper's face-image test, error after correction was still 12.5% for dark-skinned women against 1.0% for light-skinned women.
- It needs a labelled validation sample that represents the population and is large enough for the groups of concern. The guarantee for a subgroup weakens as the subgroup's share of the sample shrinks, so rare groups are protected least.
- Corrections are fitted on the small audit sample, so too many rounds or an over-flexible auditor will start fitting noise and can make performance worse on new data. The mcboost documentation says its calibration 'tends to overfit' and suggests a smaller step size or fewer iterations.
- The method corrects errors against the labels in the audit sample, so it cannot repair bias already built into those labels, such as past decisions recorded as outcomes. A fairer-looking score then still reproduces the earlier pattern.
Resources
Research Papers
mcboost: Multi-Calibration Boosting for R
Given the increasing usage of automated prediction systems in the context of high-stakes decisions, a growing body of research focuses on methods for detecting and mitigating biases in algorithmic decision-making.One important framework to audit for and mitigate biases in predictions is that of Multi-Calibration, introduced by Hebert-Johnson et al. (2018).The underlying fairness notion, Multi-Calibration, promotes the idea of multi-group fairness and requires calibrated predictions not only for marginal populations, but also for subpopulations that may be defined by complex intersections of many attributes.A simpler variant of Multi-Calibration, referred to as Multi-Accuracy, requires unbiased predictions for large collections of subpopulations.Hebert-Johnson et al. (2018) proposed a boosting-style algorithm for learning multi-calibrated predictors.Kim et al. (2019) demonstrated how to turn this algorithm into a post-processing strategy to achieve multi-accuracy, demonstrating empirical effectiveness across various domains.This package provides a stable implementation of the multi-calibration algorithm, called MCBoost.In contrast to other Fair ML approaches, MC-Boost does not harm the overall utility of a prediction model, but rather aims at improving calibration and accuracy for large sets of subpopulations post-training.MCBoost comes with strong theoretical guarantees, which have been explored formally in Hebert-Johnson et al.
Multiaccuracy
Prediction systems are successfully deployed in applications ranging from disease diagnosis, to predicting credit worthiness, to image recognition. Even when the overall accuracy is high, these systems may exhibit systematic biases that harm specific subpopulations; such biases may arise inadvertently due to underrepresentation in the data used to train a machine-learning model, or as the result of intentional malicious discrimination. We develop a rigorous framework of *multiaccuracy* auditing and post-processing to ensure accurate predictions across *identifiable subgroups*. Our algorithm, MULTIACCURACY-BOOST, works in any setting where we have black-box access to a predictor and a relatively small set of labeled data for auditing; importantly, this black-box framework allows for improved fairness and accountability of predictions, even when the predictor is minimally transparent. We prove that MULTIACCURACY-BOOST converges efficiently and show that if the initial model is accurate on an identifiable subgroup, then the post-processed model will be also. We experimentally demonstrate the effectiveness of the approach to improve the accuracy among minority subgroups in diverse applications (image classification, finance, population health). Interestingly, MULTIACCURACY-BOOST can improve subpopulation accuracy (e.g. for "black women") even when the sensitive features (e.g. "race", "gender") are not given to the algorithm explicitly.
Calibration for the (Computationally-Identifiable) Masses
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.
Software Packages
Documentations
mcboost: Multi-Calibration Boosting for R (package documentation)
Implements Multi-Calibration Boosting (2018) <https://proceedings.mlr.press/v80/hebert-johnson18a.html> and Multi-Accuracy Boosting (2019) <arXiv:1805.12317> for the multi-calibration of a machine learning models prediction. MCBoost' updates predictions for sub-groups in an iterative fashion in order to mitigate biases like poor calibration or large accuracy differences across subgroups. Multi-Calibration works best in scenarios where the underlying data & labels are unbiased, but resulting models are. This is often the case, e.g. when an algorithm fits a majority population while ignoring or under-fitting minority populations.