Meta Fair Classifier

Description

The meta fair classifier is a general-purpose training algorithm for fair classification. The user chooses a fairness metric, such as the statistical rate (each group's share of favourable outcomes) or the false discovery rate (the share of positive predictions that are wrong), and a target between 0 and 1 for how close to equal the groups' values must be. From labelled training examples that include the protected attribute, the algorithm fits a model of the data, solves a series of constrained optimisation problems, and returns the most accurate classifier that meets the target. It trains a new classifier rather than adjusting an existing one. The original paper proves the result comes close to the best classifier that meets the target, up to the error in the fitted model of the data.

Example Use Cases

Fairness

A recruitment platform trains a shortlisting classifier with a statistical-rate target of 0.8, so that the selection rate for any gender group is at least 80% of the highest group's rate, and measures how much accuracy it gives up to reach that target.

A probation service building a recidivism risk classifier constrains false discovery rate parity, so that among people flagged as high risk, the share who do not go on to reoffend is similar across racial groups.

A hospital building a classifier for heart conditions constrains false discovery rate parity, so that false alarms, which send people for unnecessary medical attention, do not fall more heavily on women than on men.

Limitations

It is a self-contained learner, not a wrapper. It cannot adjust a model you already run, and the paper lists extension to other classifier types as future work. It also needs the protected attribute for every training example, and teams must be willing to replace their model with the one it produces.
The AIF360 implementation supports only statistical rate and false discovery rate, and raises an error for any other metric. The paper's framework covers more, but other criteria, such as equalised odds, mean extra implementation work or choosing a different technique.
The guarantee depends on how well the fitted model of the data matches the real distribution, and the paper's bound carries error terms for any mismatch. In its experiments, a simple fitted model gave lower accuracy than competing methods at the same fairness level.
Results can vary from run to run. In a credit-scoring comparison, repeated runs sometimes missed the fairness target, which made averaging the results unusable.
The method solves a separate constrained optimisation problem for each value on a grid set by the tolerance, and the number multiplies for each additional fairness metric constrained at once. Combining several criteria quickly makes training slow.

Resources

Research Papers

Classification with Fairness Constraints

L. Elisa Celis et al.•Jan 29, 2019

Developing classification algorithms that are fair with respect to sensitive attributes of the data is an important problem due to the increased deployment of classification algorithms in societal contexts. Several recent works have focused on studying classification with respect to specific fairness metrics, modeled the corresponding fair classification problem as constrained optimization problems, and developed tailored algorithms to solve them. Despite this, there still remain important metrics for which there are no fair classifiers with theoretical guarantees; primarily because the resulting optimization problem is non-convex. The main contribution of this paper is a meta-algorithm for classification that can take as input a general class of fairness constraints with respect to multiple non-disjoint and multi-valued sensitive attributes, and which comes with provable guarantees. In particular, our algorithm can handle non-convex "linear fractional" constraints (which includes fairness constraints such as predictive parity) for which no prior algorithm was known. Key to our results is an algorithm for a family of classification problems with convex constraints along with a reduction from classification problems with linear fractional constraints to this family. Empirically, we observe that our algorithm is fast, can achieve near-perfect fairness with respect to various fairness metrics, and the loss in accuracy due to the imposed fairness constraints is often small.

Software Packages

AIF360

Aug 22, 2018

A comprehensive set of fairness metrics for datasets and machine learning models, explanations for these metrics, and algorithms to mitigate bias in datasets and models.

holisticai

Jan 26, 2023

This is an open-source tool to assess and improve the trustworthiness of AI systems.

Tutorials

AIF360 demo: Meta Fair Classifier (notebook)

A comprehensive set of fairness metrics for datasets and machine learning models, explanations for these metrics, and algorithms to mitigate bias in datasets and models. - Trusted-AI/AIF360

Documentations

ρ-Fair Method — holisticai documentation

Holisticai Developers•Jan 1, 2024

aif360.algorithms.inprocessing — aif360 0.1.0 documentation

Aif360 Developers•Jan 1, 2018

Related Techniques

Name	Description	Assurance Goals
Fair Bayesian Optimisation	Fair Bayesian Optimisation (FairBO) tunes a model's hyperparameters, the settings fixed before training such as tree depth or regularisation strength, to find the most accurate model that also meets a fairness threshold chosen in advance, for example a difference in positive-prediction rates between two groups of at most 0.05. Because it only changes hyperparameters and retrains, it treats the model and its training pipeline as a black box, so it works with any model class and any group fairness metric, can enforce several metrics at once, and returns a model that need not use the protected attribute at prediction time (the attribute is still needed to measure fairness during the search). Each step trains a model with a candidate configuration and measures its validation accuracy and fairness, and two Gaussian processes (statistical models that predict a value together with its uncertainty) are fitted to the results of the configurations tried so far. The next configuration is the one with the highest expected improvement in accuracy weighted by the predicted probability of meeting the fairness constraint; until a fair configuration has been found, the search maximises that probability alone. The output is the best configuration that met the constraint, with its accuracy and fairness measurements.	Fairness
Exponentiated Gradient Reduction	Exponentiated gradient reduction trains a classifier that satisfies a chosen fairness criterion, such as demographic parity (equal rates of favourable outcomes across groups) or equalised odds (equal error rates across groups), while keeping error as low as it can. It reduces the problem to a series of ordinary training runs on the same data, each with different example weights. Between runs, a multiplicative update, the exponentiated gradient step, increases the weight on fairness constraints that are currently violated, and a standard classifier is retrained on the reweighted data. The output is a randomised classifier, a weighted mix of the classifiers trained along the way, with one drawn at random for each prediction. The original paper proves that its error and its fairness violation come close to the best achievable, up to a tolerance the user sets and the statistical error from a finite training set.	Fairness
Adversarial Debiasing	Adversarial debiasing trains a model so that its predictions carry as little information as possible about a protected attribute, such as sex or race. Two neural networks are trained together, a predictor that learns the main task, such as estimating default risk, and an adversary that tries to guess each person's protected attribute from the predictor's output. The predictor is updated to stay accurate while making the adversary's guesses fail, which pushes its outputs towards equal rates of favourable outcomes across groups (demographic parity). If the adversary is also given the true label, the predictor is penalised only for information about the attribute beyond what the label carries, which targets equal error rates across groups (equalised odds). The method needs the protected attribute for every training example and a model trained by gradient descent.	Fairness