Entities that use algorithmic systems in traditional civil rights domains like housing, employment, and credit should have a duty to search for and implement less discriminatory algorithms (LDAs). Why? Work in computer science has established that, contrary to conventional wisdom, for a given prediction problem there are almost always multiple possible models with equivalent performance—a phenomenon termed model multiplicity. Critically for our purposes, different models of equivalent performance can produce different predictions for the same individual, and, in aggregate, exhibit different levels of impacts across demographic groups. As a result, when an algorithmic system displays a disparate impact, model multiplicity suggests that developers may be able to discover an alternative model that performs equally well, but has less discriminatory impact. Indeed, the promise of model multiplicity is that an equally accurate, but less discriminatory alternative algorithm almost always exists. But without dedicated exploration, it is unlikely developers will discover potential LDAs.

Model multiplicity has profound ramifications for the legal response to discriminatory algorithms. Under disparate impact doctrine, it makes little sense to say that a given algorithmic system used by an employer, creditor, or housing provider is either “justified” or “necessary” if an equally accurate model that exhibits less disparate effect is available and possible to discover with reasonable effort. Indeed, the overarching purpose of our civil rights laws is to remove precisely these arbitrary barriers to full participation in the nation’s economic life, particularly for marginalized racial groups. As a result, the law should place a duty of a reasonable search for LDAs on entities that develop and deploy predictive models in covered civil rights domains. The law should recognize this duty in at least two specific ways. First, under disparate impact doctrine, a defendant’s burden of justifying a model with discriminatory effects should be recognized to include showing that it made a reasonable search for LDAs before implementing the model. Second, new regulatory frameworks for the governance of algorithms should include a requirement that entities search for and implement LDAs as part of the model building process.

While algorithmic fairness is a thriving area of research, in practice, mitigating issues of bias often gets reduced to enforcing an arbitrarily chosen fairness metric, either by enforcing fairness constraints during the optimization step, post-processing model outputs, or by manipulating the training data. Recent work has called on the ML community to take a more holistic approach to tackle fairness issues by systematically investigating the many design choices made through the ML pipeline, and identifying interventions that target theissue’s root cause, as opposed toits symptoms.While we share the conviction that this pipeline-based approach is the most appropriate for combating algorithmic unfairness on the ground, we believe there are currently very few methods of operationalizing this approach in practice. Drawing on our experience as educators and practitioners, we first demonstrate that without clear guidelines and toolkits, even individuals with specialized ML knowledge find it challenging to hypothesize how various design choices influence model behavior. We then consult the fair-ML literature to understand the progress to date toward operationalizing the pipeline-aware approach - we systematically collect and organize the prior work that attempts to detect, measure, and mitigate various sources of unfairness through the ML pipeline. We utilize this extensive categorization of previous contributions to sketch a research agenda for the community. We hope this work serves as the stepping stone toward a more comprehensive set of resources for ML researchers, practitioners, and students interested in exploring, designing, and testing pipelineoriented approaches to algorithmic fairness.

The vast majority of techniques to train fair models require access to the protected attribute (e.g., race, gender), either at train time or in production. However, in many important applications this protected attribute is largely unavailable. In this paper, we develop methods for measuring and reducing fairness violations in a setting with limited access to protected attribute labels. Specifically, we assume access to protected attribute labels on a small subset of the dataset of interest, but only probabilistic estimates of protected attribute labels (e.g., via Bayesian Improved Surname Geocoding) for the rest of the dataset. With this setting in mind, we propose a method to estimate bounds on common fairness metrics for an existing model, as well as a method for training a model to limit fairness violations by solving a constrained non-convex optimization problem. Unlike similar existing approaches, our methods take advantage of contextual information – specifically, the relationships between a model’s predictions and the probabilistic prediction of protected attributes, given the true protected attribute, and vice versa – to provide tighter bounds on the true disparity. We provide an empirical illustration of our methods using voting data. First, we show our measurement method can bound the true disparity up to 5.5x tighter than previous methods in these applications. Then, we demonstrate that our training technique effectively reduces disparity while incurring lesser fairness-accuracy trade-offs than other fair optimization methods with limited access to protected attributes.

Recent work has shown that models trained to the same objective, and which achieve similar measures of accuracy on consistent test data, may nonetheless behave very differently on individual predictions. This inconsistency is undesirable in high-stakes contexts, such as medical diagnosis and finance. We show that this inconsistent behavior extends beyond predictions to feature attributions, which may likewise have negative implications for the intelligibility of a model, and one’s ability to find recourse for subjects. We then introduce selective ensembles to mitigate such inconsistencies by applying hypothesis testing to the predictions of a set of models trained using randomly-selected starting conditions; importantly, selective ensembles can abstain in cases where a consistent outcome cannot be achieved up to a specified confidence level. We prove that that prediction disagreement between selective ensembles is bounded, and empirically demonstrate that selective ensembles achieve consistent predictions and feature attributions while maintaining low abstention rates. On several benchmark datasets, selective ensembles reach zero inconsistently predicted points, with abstention rates as low 1.5%.

Artificial Intelligence (AI) systems now influence decisions impacting every aspect of people’s lives, from the news articles they read, to whether or not they receive a loan. While the use of AI may lead to great accuracy and efficiency in the making of these important decisions, recent news and research reports have shown that AI models can act unfairly - from exhibiting gender bias in hiring models, to racial bias in recidivism prediction systems.

This thesis explores new methods for understanding and mitigating fairness issues in AI through considering how choices made throughout the process of creating an AI system—i.e., the modeling pipeline—impacts fairness behavior. First, I will show how considering a model’s end-to-end pipeline allows us to expand our understanding of unfair model behavior. In particular, my work introduces a connection between AI system stability and fairness by demonstrating how instability in certain parts of the modeling pipeline, namely the learning rule, can lead to unfairness by having important decisions rely on arbitrary modeling choices.

Secondly, I will discuss how considering ML pipelines can help us expand our toolbox of bias mitigation techniques. In a case study investigating equity with respect to income in tax auditing practices, I will demonstrate how interventions made along the AI creation pipeline—even those not related to fairness on their face—can not only be effective for increasing fairness, but can often reduce tradeoffs between predictive utility and fairness.

Finally, I will close with an overview of the benefits and dangers of the flexibility that the AI modeling pipeline affords practitioners in the creation of their models, including a discussion of the the legal repercussions of this flexibility, which I call model multiplicity.