Source-linked AI summary
Learning Adversarially Fair and Transferable Representations
David Madras, Elliot Creager, Toniann Pitassi, Richard Zemel
TL;DR
Third-party vendors may use useful representations for unfair predictions, creating a need for representations that protect fairness under unknown downstream objectives. The paper develops adversarial objectives tied to group-fairness metrics and finds fair, useful predictions, including on transferred tasks, while identifying open questions about objective design and transfer scope.
Problem
Unknown or fairness-indifferent downstream vendors may produce unfair predictions, while evidence remains limited on representations that preserve utility and fairness across new tasks.
Method
The paper learns representations with adversaries whose objectives are designed for demographic parity, equalized odds, and equal opportunity, and analyzes their worst-case fairness guarantees.
Results
The learned representations support fair and useful predictors, including fair transfer to unseen tasks; in transfer experiments, LAFTR improved ΔEO by an average of ∼20% over the non-transfer baseline with only marginally worse accuracy than other models.
Takeaways & Limitations
Fairness can be enforced at the representation level so that predictors learned by third parties, including on new tasks, exhibit fairness improvements without explicit fairness constraints.
Takeaways & Limitations
Open questions remain about comparing approaches, choosing and stabilizing adversarial losses, and determining when and between which tasks fair transfer succeeds.
Abstract
from arXiv · showhide
In this paper, we advocate for representation learning as the key to mitigating unfair prediction outcomes downstream. Motivated by a scenario where learned representations are used by third parties with unknown objectives, we propose and explore adversarial representation learning as a natural method of ensuring those parties act fairly. We connect group fairness (demographic parity, equalized odds, and equal opportunity) to different adversarial objectives. Through worst-case theoretical guarantees and experimental validation, we show that the choice of this objective is crucial to fair prediction. Furthermore, we present the first in-depth experimental demonstration of fair transfer learning and demonstrate empirically that our learned representations admit fair predictions on new tasks while maintaining utility, an essential goal of fair representation learning.
1. Introduction
The paper reframes the data owner’s representation choice as an adversarial representation-learning problem to help ensure downstream predictors remain fair while retaining utility.
- Motivation: Separating data ownership from prediction creates a risk that accuracy-driven vendors produce unfair predictions from historically discriminatory data.The paper highlights online advertising as an owner-vendor setting in which predictive incentives may conflict with fairness.
- Motivation: The data owner must choose representations that constrain downstream unfairness without sacrificing relatively high predictive utility.Providing more data may maximize utility, but does not ensure that vendors learn fair predictors.
- Approach: The paper frames representation selection as learning with an adversary that criticizes potentially unfair solutions.This shifts fairness protection toward the representation supplied to third-party prediction vendors.
- Contributions: The method connects demographic parity, equalized odds, and equal opportunity to adversarial objectives with worst-case guarantees for downstream classifiers.The paper also distinguishes its emphasis on adversarial criticism of representations from prior adversarial-fairness approaches.
- Contributions: Experiments evaluate both fair classification and fair transfer, including whether learned representations support fair predictors on unseen tasks without explicit fairness constraints.The paper presents these experiments as validation of fair and transferable representations.
2. Background
The background introduces group fairness criteria and adversarial learning, while noting that fairness constraints can conflict with calibration and require utility-fairness tradeoffs.
- Fairness: Fair classification seeks accurate predictions with respect to labels Y while avoiding bias with respect to sensitive attributes A.The setup uses binary labels, sensitive attributes, and predictions.
- Fairness: Demographic parity requires equal positive-outcome rates across groups: P( ˆY = 1|A = 0) = P( ˆY = 1|A = 1).When group base rates differ, equalized odds and equal opportunity condition fairness on the ground truth Y.
- Fairness: Fairness constraints can conflict with well-calibrated classifiers, motivating relaxed objectives that trade utility against fairness through hyperparameters.The usual utility target is classification accuracy.
- Adversarial Learning: Adversarial learning trains two competing models, such as a generator and discriminator, through max-min optimization.The discriminator distinguishes real from generated data while the generator attempts to fool it.
3. Related Work
Related work spans group and individual fairness, dataset processing, fair representations, domain adaptation, and adversarial classifiers; this paper instead targets fairness at the learned representation.
- Fairness Research: Earlier fairness research addressed group fairness, individual fairness, and their intersection across multiple group identities.The cited work includes definitions and methods combining elements of group and individual fairness.
- Fair Representations: Prior methods explored fairness through pre-processing and post-processing datasets, separate stakeholder frameworks, and maximum mean discrepancy for representation learning.These approaches differ in where fairness constraints are applied.
- Adversarial Methods: Adversarial representation learning for domain adaptation provides a related framework because multiple distinct data distributions must remain expressively modeled.The paper identifies this as a conceptual connection to fair representation learning.
- Adversarial Fairness: Closest prior work adversarially learned classifiers for demographic parity, whereas this paper lets the adversary access the learned representation directly.The stated goal is fair and transferable representations that admit fair classification outcomes.
4. Adversarially Fair Representations
The model learns representations that reconstruct inputs, support target prediction, and hide sensitive attributes from an adversary; its objectives are tailored to fairness criteria and downstream vendor behavior.
- Generalized Model: The generalized model learns Z to reconstruct X, classify Y, and protect sensitive attribute A from an adversary.Reconstruction or classification can be omitted by setting corresponding hyperparameters to zero.
- Generalized Model: The encoder f produces Z, while classifier g predicts Y, adversary h predicts A, and optional decoder k reconstructs X using Z and A.For equalized odds or equal opportunity, the adversary also receives Y.
- Optimization: The adversary maximizes its fairness objective while the encoder, decoder, and classifier minimize classification loss, reconstruction error, and the adversary objective.Training alternates gradient descent for f, g, k with gradient ascent for h.
- Design Choices: The hyperparameters α, β, γ balance utility, input reconstruction, and fairness, and the model is named LAFTR because of its focus on fair transfer learning.The name expands to Learned Adversarially Fair and Transferable Representations.
- Fairness Objectives: The adversarial objective is modified by fairness desideratum, using group-based discrepancies for demographic parity and equalized odds, with equal opportunity summing only terms for Y = 0.The objective is designed to align adversarial training with the selected group-fairness metric.
- Vendor Model: The data owner simulates an indifferent vendor through the classifier and an adversarial vendor through the adversary, so defeating a strong adversary bounds downstream unfairness.The paper motivates strong adversaries because downstream classifiers trained naively on Z should not exceed the adversary’s unfairness.
5. Theoretical Properties
The paper derives adversarial objectives whose optimal values upper-bound demographic parity and equalized odds unfairness for downstream classifiers. It also explains how these guarantees motivate adversarial representation learning and why the chosen loss must track test discrepancy.
- Metric-specific adversarial objectives: Adversarial upper bounds connect suitable objectives to demographic parity, equalized odds, and equal opportunity.The objective is selected according to the desired fairness metric.
- Demographic parity: Demographic parity distance is zero exactly when classifier outcomes are independent of group membership.The representation distributions are conditioned on A = 0 and A = 1.
- Demographic parity: For demographic parity, the optimal adversary’s objective upper-bounds the demographic parity distance of every classifier learnable from the representation.The bound is expressed as Adv(h∗) ≥ ∆DP(g).
- Equalized odds: For equalized odds, an optimal adversary receiving both representation and label upper-bounds the classifier’s equalized odds distance.The distance combines absolute differences in false-positive and false-negative rates.
- Proof interpretation: The proofs construct adversaries from the joint distribution of group, label, and classifier output, without requiring direct access to the representation.With access restricted to these outputs, the adversarial objective is equivalent to adding the corresponding fairness distance to the classification objective.
- Objective design: The proposed group-normalized ℓ1 objective is intended to measure test discrepancy more naturally than cross-entropy, which can fail on imbalanced datasets.The paper presents group-normalized ℓ1 as the theoretically motivated relaxation, although gradient optimization can be difficult.
6. Experiments
Experiments evaluate LAFTR for fair classification and fair transfer learning using frozen learned representations and unconstrained downstream classifiers. LAFTR achieves strong fairness-accuracy tradeoffs and substantially improves transfer fairness while retaining comparable utility.
- Evaluation procedure: LAFTR evaluates representations by freezing the encoder and training an unconstrained classifier on top of its outputs.The composed classifier is assessed for accuracy and fairness on held-out data, including new transfer tasks.
- Fair classification: On Adult fair classification, the objective matched to each target fairness metric tends to achieve the best trade-off, especially at small unfairness.The preferred region combines high accuracy with low fairness violation.
- Fair classification: At low unfairness, the criterion-specific LAFTR model obtains the highest test accuracy, although performance is less consistent elsewhere because adversarial training can be unstable.This supports the potential of matching adversarial objectives to fairness criteria.
- Fair transfer learning: Across 10 Health transfer tasks, LAFTR improves ∆EO by an average of ∼20% relative to the non-transfer baseline while suffering only a marginally larger accuracy loss than other models.The comparison uses relative changes in error and ∆EO against a high-accuracy model trained directly from the data.
- Fair transfer learning: Fairly regularized and output-only adversarial MLPs do not match LAFTR’s transfer fairness, highlighting the value of an adversary that observes representations.The regularized MLP performs similarly to the unconstrained transfer baseline, and the output-only adversary has similar results.
- Fair transfer learning: LAFTR provides the fairest predictions on 7 of 10 transfer tasks and is never far behind the fairest model on each task.The unfair baseline is fairest on one task, while LAFTR remains relatively fair when that baseline performs best and avoids its large fairness failures elsewhere.
7. Conclusion
The paper develops adversarially fair representations with theoretical grounding and demonstrates that they support both useful prediction and fair transfer to new tasks. It also identifies unresolved questions about comparing approaches, adversarial losses, and when transfer fairness succeeds.
- 7. Conclusion: Adversarially fair representations receive theoretical grounding through novel objectives tied to commonly used group-fairness metrics.The objectives are designed to guarantee performance on those metrics.
- 7. Conclusion: Experiments show that an adversary can produce representations supporting fair and useful predictors.The adversary operates on the intermediate representation.
- 7. Conclusion: Fair transfer experiments show utility transfer to new tasks alongside fairness improvements.The conclusion presents this as evidence that the learned representations transfer beyond their original task.
- 7. Conclusion: Open problems include in-depth comparisons of adversarial and non-adversarial approaches, including MMD, to clarify their pros and cons.The paper specifically calls for careful comparative analysis.
- 7. Conclusion: Open questions remain about optimal adversarial losses and which task pairs and situations are most likely to support fair transfer.The authors connect better loss-function understanding with stabilizing adversarial training.
A. Understanding cross entropy loss in fair adversarial training
The adversary is intended to measure statistical discrepancy between sensitive groups, but standard cross-entropy can fail to do so on imbalanced data. A group-normalized objective instead directly optimizes the relevant test discrepancy.
- A. Understanding cross entropy loss in fair adversarial training: An approximately optimal adversary makes its objective close to the statistical distance between Z0 and Z1, regularizing representations by that distance.The adversary functions as a test-discrepancy estimator for a chosen representation.
- A. Understanding cross entropy loss in fair adversarial training: Cross-entropy does not calculate the test discrepancy of a given adversary, so its statistical-distance interpretation breaks down.The paper explicitly identifies this issue for the cross-entropy objective.
- A. Understanding cross entropy loss in fair adversarial training: On imbalanced datasets, cross-entropy favors correctly predicting the majority class instead of finding a larger test discrepancy.The standard loss averages over the whole dataset, whereas test discrepancy sums conditional expectations across groups.
- A. Understanding cross entropy loss in fair adversarial training: In the toy example, cross-entropy classifies every Z as A-hat = 0 and achieves test discrepancy d_h(Z0, Z1) = 0.This illustrates the failure mode caused by dataset imbalance.
- A. Understanding cross entropy loss in fair adversarial training: Directly optimizing the suggested test discrepancy yields approximately 0.368 in the toy example, outperforming cross-entropy for discrepancy optimization.The comparison is stated against the group-normalized ℓ1 objective.
B. Training Details
Training uses small single-hidden-layer neural networks, dataset-specific latent dimensions, absolute-error objectives, and Adam optimization. Fairness-accuracy curves sweep the fairness coefficient and evaluate checkpoints using repeated validation and test classifiers.
- B. Training Details: Training uses absolute error for both classification and adversarial losses, with leaky ReLU networks and Adam at learning rate 0.001.The minibatch size is 64, with one update step per minibatch for the encoder-classifier and discriminator.
- B. Training Details: Fairness-accuracy curves sweep γ from 0.1 to 4 and select checkpoints using validation error plus Δ across seven classifier seeds.Seven additional classifiers are then trained on an unseen test set.
- B. Training Details: Transfer-learning experiments use γ = 1 for models requiring a fair regularization coefficient.This setting applies specifically to the transfer experiment.