Don’t Throw it Away! The Utility of Unlabeled Data in Fair Decision Making

Abstract

Decision making algorithms, in practice, are often trained on data that exhibits a variety of biases. Decision-makers often aim to take decisions based on some ground-truth target that is assumed or expected to be unbiased, i.e., equally distributed across socially salient groups. In many practical settings, the ground-truth cannot be directly observed, and instead, we have to rely on a biased proxy measure of the ground-truth, i.e., biased labels, in the data. In addition, data is often selectively labeled, i.e., even the biased labels are only observed for a small fraction of the data that received a positive decision. To overcome label and selection biases, recent work proposes to learn stochastic, exploring decision policies via i) online training of new policies at each time-step and ii) enforcing fairness as a constraint on performance. However, the existing approach uses only labeled data, disregarding a large amount of unlabeled data, and thereby suffers from high instability and variance in the learned decision policies at different times. In this paper, we propose a novel method based on a variational autoencoder for practical fair decision-making. Our method learns an unbiased data representation leveraging both labeled and unlabeled data and uses the representations to learn a policy in an online process. Using synthetic data, we empirically validate that our method converges to the optimal (fair) policy according to the ground-truth with low variance. In real-world experiments, we further show that our training approach not only offers a more stable learning process but also yields policies with higher fairness as well as utility than previous approaches.

Fairness and Sequential Decision Making: Limits, Lessons, and Opportunities

“…In this section, we examine common interventions, including data augmentation, reward modification, optimization formulation, and system integration. For predictive models one of the simplest interventions is to collect or generate more data (data augmentation), under the assumption that as more samples are acquired the training set will improve its approximation of the true distribution and the model will learn a more accurate, representative function [176,191,204,227,241,252]. This practice works well if data deficiency is the only reason for poor performance.…”

An Anti-Subordination Approach to Fair Classification

“…However, their method requires non-trivial parametric assumptions on the feature distribution. Rateike et al [31] develop an online process that first learns an unbiased representation of the data and then trains an online classifier over the learned representation space. This approach and the above papers, however, do not employ any constraint on false positives which can lead to low utility in certain iterations when sufficient information for learning is unavailable (see Section 5 for empirical comparison against these methods).…”

“…Baselines. We compare our approach to the following baselines: (a) KILBERTUS ET AL [22], which uses stochastic classifiers to assign a non-zero exploration probability to every sample; (b) YANG ET AL [38], which employs a bandit-type approach, first determining the likelihoods using a logistic model and then adjusting classifier thresholds to incorporate gathered information; (c) RATEIKE ET AL [31], which learns an unbiased representation of the data using which an online decisionmaking model is trained; (d) OPT-OFFLINE, i.e. the ideal (unattainable in partial feedback setting) classifier trained using i.i.d.…”

“…• RATEIKE ET AL. The algorithm from the Rateike et al [31] paper is implemented the code provided by the authors 4 . We execute this baseline using the initial data and batch size configuration specified in Section 5.…”

I Prefer not to Say: Operationalizing Fair and User-guided Data Minimization

“…Classification models that can handle missing data have been studied in the previous literature with the goal of minimizing costs or increasing performance [44,2], obtaining uncertainty estimates [20], or fulfilling classical fairness notions [45,19,40,11]. In a related line of work, classification with noisy [12] or missing labels [22,35] has been investigated, where the missingness is often a result of selection bias. The setting considered in this work is different in the sense that we are not concerned with fulfilling a fairness notion with respect to sensitive information, but between subjects with and without optional information.…”