Multiaccuracy: Black-Box Post-Processing for Fairness in Classification

If the initial model is accurate on an identifiable subgroup, then the post-processed model will also be accurate on that subgroup.

“Importantly, if the initial model is accurate on an identifiable subgroup, then the post-processed model will be also.”

The paper proves convergence guarantees for Multiaccuracy Boost.

“We prove convergence guarantees for Multiaccuracy Boost and show that post-processing for multiaccuracy may actually improve the overall classification accuracy.”

Multiaccuracy Boost can improve subpopulation accuracy even when sensitive features such as race and gender are not known to the algorithm.

“Multiaccuracy Boost can improve subpopulation accuracy (e.g. for”

Multiaccuracy Boost improves overall accuracy at each step.

“We prove guarantees on the convergence rate of the algorithm and show that it improves overall accuracy at each step.”

Enforcing multiaccuracy actually improves overall accuracy by identifying subpopulations where the initial models systematically erred.

“As suggested by the theory, enforcing multiaccuracy actually improves the overall accuracy, by identifying subpopulations where the initial models systematically erred”

The do-no-harm property guarantees that on any subpopulation S that A audits, the classification error cannot increase significantly from f0 to the post-processed classifier.

“In a nutshell, the property guarantees that on any subpopulation S ⊆ X that A audits, the classification error cannot increase significantly from f 0 to the post-processed classifier.”

• AssumptionThe authors state this as an assumption: The auditing algorithm A can identify a very rich class of overlapping sets.

Post-processing does not significantly affect performance on groups where accuracy was already high.

“further, post-processing does not significantly affect performance on groups where accuracy was already high.”

The initial inception-resnet-v1 gender classification model trained on CelebA achieved 98.4% test accuracy on CelebA for binary gender classification.

“The resulting test accuracy on CelebA for binary gender classification is 98.4%.”

In an experimental case study based on Gender Shades, after post-processing, accuracy improves across minority groups even though minority-status is not explicitly given to the post-processing algorithm as a feature.

“After post-processing, the accuracy improves across these minority groups, even though minority-status is not explicitly given to the post-processing algorithm as a feature.”

For any subpopulation S with Pr[x ∈ S] ≥ γ, if A audits χ_S and its negation and f is the output of Algorithm 1 with α ≤ βγ, then er_S(f; y) ≤ 3 • er_S(f0; y) + 4β.

“Then the classification error of f on the subset S is bounded as er S (f ; y) ≤ 3 • er S (f 0 ; y) + 4β .”

Post-processing for multiaccuracy may actually improve overall classification accuracy.

“We prove convergence guarantees for Multiaccuracy Boost and show that post-processing for multiaccuracy may actually improve the overall classification accuracy.”

Multiaccuracy post-processing improves generalization error before overfitting.

“we show empirically that multiaccuracy post-processing improves the generalization error before overfitting.”

The multiaccuracy auditor can be used to produce examples of inputs where the predictor is erring significantly.

“Specifically, the multiaccuracy auditor can be used to produce examples of inputs where the predictor is erring significantly.”

When Multiaccuracy Boost terminates, f_T will be (C, α)-multiaccurate.

“From the stopping condition, it is clear that when the algorithm terminates, f T will be (C, α)-multiaccurate.”

Multiaccuracy Boost outputs a hypothesis f that is (C, α)-multiaccurate.

“outputs a hypothesis f : X → [0, 1] that is (C, α)-multiaccurate.”

Multiaccuracy Boost exhibits a do-no-harm guarantee: if f0 has low classification error on some subpopulation S identified by A, then the resulting classification error on S cannot increase significantly.

“Importantly, we adapt these algorithms so that Multiaccuracy Boost exhibits what we call the”

Empirically, classification accuracy on the held-out test set tends to improve over D after multiaccuracy post-processing.

“empirically, we observe that classification accuracy (on held-out test set) tends to improve over D after multiaccuracy post-processing.”

If there is a function in C that correlates well with the label function on a significant subpopulation S, then multiaccuracy translates into a guarantee on classification accuracy on S.

“That is, if there is a function in C that correlates well with the label function on a significant subpopulation S, then multi-accuracy translates into a guarantee on the classification accuracy on this subpopulation.”

The least active images in the first round of auditing are men in suits.

“The least active images (second row) are men in suits”

The only way the authors were able to report these accuracy disparities was by having access to a rich data set where gender and race were labeled.

“Importantly, the only way we were able to report these accuracy disparities was by having access to a rich data set where gender and race were labeled.”

Models trained in good faith on unbalanced data may exhibit significant biases on minority populations.

“we've demonstrated that models trained in good faith on unbalanced data may exhibit significant biases on the minority populations.”

In the first round of auditing, the three highest-scoring images are all women, both black and white.

“In the first round of auditing, the three highest-scoring images (top row) are all women, both black and white.”

As Table 3 shows, Multiaccuracy Boost significantly lowers the classification error in the old female population.

“As Table 3 shows, Multiaccuracy Boost significantly lowers the classification error in the old female population.”

In our experiments, standard supervised learning optimizes for overall performance, leading to settings where certain subpopulations incur substantially worse error rates.

“In our experiments, we observe that standard supervised learning optimizes for overall performance, leading to settings where certain subpopulations incur substantially worse error rates.”

Machine learning predictors often have systematic biases that harm specific subgroups, especially subgroups that are minorities in the training data.

“Despite the successes of machine learning at complex tasks that involve making predictions about people, there is growing evidence that”

• cites [2]Not checked yet

  Not checked in the free run.

In the Gender Shades study, the systems were significantly less accurate on female subjects compared to males.

“the system was significantly less accurate on female subjects compared to males”

• cites [2]Not checked yet

  Not checked in the free run.

In the Gender Shades study, classification accuracy differed between dark-skinned females and light-skinned males by as much as 34%.

“classification accuracy differed between these groups by as much as 34%!”

• cites [2]Not checked yet

  Not checked in the free run.

In the Gender Shades study, the systems were significantly less accurate on dark-skinned individuals compared to light-skinned individuals.

“and on dark-skinned individuals compared to light-skinned.”

• cites [2]Not checked yet

  Not checked in the free run.

Multiaccuracy Boost works in any setting where there is black-box access to a predictor and a relatively small set of labeled data for auditing.

“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.”

If the auditor A learns an expressive enough class of functions, then the algorithm will start to overfit at some point.

“Indeed, if the auditor A learns an expressive enough class of functions, then our algorithm will start to overfit at some point”

Requiring multiaccuracy with respect to the class of all statistical tests requires learning the function y(x) exactly, which is information-theoretically impossible from a small sample.

“Requiring multiaccuracy with respect to such a C, however, requires learning the function y(x) exactly, which is informationtheoretically impossible from a small sample.”

Achieving the paper's notion of fairness need not adversely affect the utility of the classifier.

“In this sense, achieving our notion of fairness need not adversely affect the utility of the classifier.”

• AssumptionThe authors state this as an assumption: The auditing algorithm A can identify a very rich class of overlapping sets.

If Algorithm 1 is used to post-process a model that already achieves high accuracy on the validation distribution, the resulting model's accuracy should not deteriorate in significant ways.

“Thus, if we use Algorithm 1 to post-process a model that is already achieves high accuracy on the validation distribution the resulting model's accuracy should not deteriorate in significant ways”

• AssumptionThe authors state this as an assumption: The auditing algorithm A can identify a very rich class of overlapping sets.

As long as there are features in the audit set correlated with the unobserved human categories, Multiaccuracy Boost is effective at improving classification accuracy across these categories.

“As long as there are features in the audit set correlated with the (unobserved) human categories, then Multiaccuracy Boost is effective at improving the classification accuracy across these categories.”

If C is a learnable class of functions, then (C, α)-multiaccuracy guarantees accuracy on all efficiently-identifiable subpopulations.

“In practice, if we take C to be a learnable class of functions, then (C, α)-multiaccuracy guarantees accuracy on all efficiently-identifiable subpopulations.”

Partitioning the search space X based on the predictions of f0 helps ensure that the output f maintains the initial accuracy of f0.

“Partitioning the search space X based on the predictions of f 0 helps to ensure that the f we output maintains the initial accuracy of f 0”

If C is the class of depth-4 decision trees, then multiaccuracy guarantees unbiasedness on subpopulations defined by combinations of race, gender, and two other features, including overlapping groups beyond traditionally-protected populations.

“For instance, if we took C to be the class of depth-4 decision trees, then multiaccuracy guarantees unbiasedness, not just on the marginal populations defined by race and separately gender, but by the subpopulations defined by the combinations of race, gender, and two other (possibly”

Multiaccuracy auditing may serve as an effective diagnostic and interpretation tool to help developers identify systematic biases in their models.

“We demonstrate that multiaccuracy auditing may serve as an effective diagnostic and interpretation tool to help developers identify systematic biases in their models.”

Our approach to post-processing, which uses a learning algorithm as a fairness auditor, differs from [16] in important ways.

“Our approach to post-processing, which uses a learning algorithm as a fairness auditor differs from [16] in important ways.”

• cites [16]Not checked yet

  Not checked in the free run.

Suits may be a highly predictive feature of being a man according to the original classifier f0.

“suggesting that suits may be a highly predictive feature of being a man according to the original classifier, f 0 .”

The do-no-harm property guarantees that if the initial predictor f0 has small classification error on some identifiable group, then the resulting post-processed model will also have small classification error on this group.

“This property guarantees that if the initial predictor f 0 has small classification error on some identifiable group, then the resulting post-processed model will also have small classification error on this group.”

Absent any identification of the subpopulations on which the model is underperforming, collecting additional training data may not actually improve performance across the board.

“Absent any identification of the subpopulations on which the model is underperforming, collecting additional training data may not actually improve performance across the board.”

By applying multiaccuracy over real-valued functions (as opposed to boolean functions), we can get efficient algorithms that provably satisfy (C, α)-multiaccuracy for nontrivial classes C, specifically linear hypotheses.

“Additionally, by applying mutliaccuracy over real-valued functions (as opposed to boolean functions), we can get efficient algorithms that provably satisfy (C, α)-multiaccuracy for nontrivial classes C (specifically, linear hypotheses).”

Multiaccuracy allows us to protect a rich class of overlapping subpopulations without explicit knowledge of the vulnerable populations.

“In contrast, multiaccuracy allows us to protect a rich class of overlapping subpopulations without explicit knowledge of the vulnerable populations.”

It becomes important that we leverage the expressiveness and initial accuracy of f0 if we are to obtain strong performance using the multiaccuracy approach.

“Thus, it becomes important that we leverage the expressiveness (and initial accuracy) of f 0 if we are to obtain strong performance using the multiaccuracy approach.”

In the second round, after one step of Multiaccuracy Boost, the auditor seems to hone in on the dark-skinned women subpopulation as the region of bias.

“In the second round, after the classifier has been improved by one step of Multiaccuracy Boost, the auditor seems to hone in on the”

The first round of audit seems to identify gender as the axis of bias in f0.

“Overall the first round of audit seems to identify gender as the axis of bias in f 0 .”

While applicable in some settings, the decoupled classifier approach of Dwork et al. [8] may at times be untenable.

“While applicable in some settings, at times, this approach may be untenable.”

• cites [8]Not checked yet

  Not checked in the free run.

Multiaccuracy works to the extent that the auditor can identify specific subgroups where the original classifier f0 tends to make mistakes.

“Multiaccuracy works to the extent that the auditor can identify specific subgroups where the original classifier f 0 tends to make mistakes.”

Even when subgroups are identified, there are often not enough samples to train accurate classifiers on each subgroup separately.

“Even when the subgroups are identified, we often do not have enough samples to train accurate classifiers on each subgroup separately.”

• AssumptionRests on an assumption: Even when subgroups are identified, there are often not enough samples to train accurate classifiers on each subgroup separately.

The subgroup-specific models require explicit access to the sensitive features of gender and race.

“Importantly, the subgroup-specific models requires explicit access to the sensitive features of gender and race.”

In realistic settings, post-processing for multiaccuracy helps to mitigate systematic biases in predictors across sensitive subpopulations, even when the identifiers for these subpopulations are not given to the auditor explicitly.

“We show that in realistic settings, post-processing for multiaccuracy helps to mitigate systematic biases in predictors across sensitive subpopulations, even when the identifiers for these subpopulations are not given to the auditor explicitly.”

Multiaccuracy auditing and post-processing is applicable broadly in supervised learning tasks, not just in image classification applications.

“Multiaccuracy auditing and post-processing is applicable broadly in supervised learning tasks, not just in image classification applications.”

The subpopulations on which a classifier errs may be efficiently-identifiable.

“Our empirical results suggest, that the subpopulations on which a classifier errs may be efficiently-identifiable.”

Multiaccuracy auditing is limited: if the mistakes appear overly-complicated to the bounded auditor, then the auditor will not be able to identify these mistakes.

“Thus, multiaccuracy auditing is limited: if the mistakes appear overly-complicated to the bounded auditor, then the auditor will not be able to identify these mistakes.”

• AssumptionRests on an assumption: The auditor is bounded.

Requiring subgroup identification is not feasible in many applications where race or more granular categories are not given.

“This is not feasible in the many applications where, say, race or more granular categories are not given.”

• AssumptionRests on an assumption: In many applications, race or more granular categories are not given.

The power of multiaccuracy lies in the fact that in many settings, we can identify issues with f0 from a small amount of audit data.

“The power of multiaccuracy lies in the fact that in many settings, we can identify issues with f 0 from a small amount of audit data.”

Training a classifier for each subgroup, or explicitly adding subgroup accuracy into the training objective, assumes that the subgroup is already identified in the data.

“Training a classifier for each subgroup, or explicitly adding subgroup accuracy into the training objective, assumes that the subgroup is already identified in the data.”

The adult income example illustrates that multiaccuracy can help to boost the overall accuracy of a black-box predictor in a data efficient manner.

“This example illustrates that multiaccuracy can help to boost the overall accuracy of a black-box predictor in a data efficient manner.”

Our experiments improving the accuracy of a model trained on CelebA on the PPB test sets suggest a lightweight black-box alternative to more sophisticated transfer learning techniques.

“In particular, our experiments improving the accuracy of a model trained on CelebA on the PPB test sets suggests a lightweight black-box alternative to more sophisticated transfer learning techniques, which may warrant further investigation.”

Theorem 3: Let α, δ > 0 and suppose A agnostic learns a class C ⊆ [-1, 1]^X of dimension d(C). Then, using η = O(α), Algorithm 1 converges to a (C, α)-multiaccurate hypothesis fT in T = O(lD(f0;y)/α^2) iterations from m = Õ(T·d(C)+log(1/δ)/α^2) samples with probability at least 1-δ over the random samples.

“Theorem 3. Let α, δ > 0 and suppose A agnostic learns a class C ⊆ [-1, 1] X of dimension d(C).Then, using η = O(α), Algorithm 1 converges to a (C, α)-multiaccurate hypothesis f T in T = O ℓ D (f 0 ;y) α 2 iterations from m = Õ T • d(C)+log(1/δ ) α 2 samples with probability at least 1 -δ over the random samples.”

• AssumptionThe authors state this as an assumption: A agnostic learns a class C ⊆ [-1, 1]^X of dimension d(C).