As data becomes the fuel driving technological and economic growth, a fundamental challenge is how to fairly quantify the value of data in algorithmic predictions and decisions. For example, Gov. Newsom recently proposed "data dividend" whereby consumers are compensated by companies for the data that they generate. In this work, we develop a principled framework to address data valuation in the context of supervised machine learning. Given a learning algorithm trained on n data points to produce a predictor, we study data Shapley as an equitable metric to quantify the value of each training datum to the predictor performance. Data Shapley uniquely satisfies several natural properties of equitable data valuation. We develop Monte Carlo and gradient-based methods to efficiently estimate data Shapley values in practical settings where complex learning algorithms, including neural networks, are trained on large datasets. In addition to being equitable, our experiments across biomedical, image and synthetic data demonstrate that data Shapley has several other benefits: 1) it gives actionable insights on what types of data benefit or harm the prediction model; 2) weighting training data by Shapley value improves domain adaptation. This is joint work with Amirata Ghorbani.

We consider the problem of learning representations that achieve group and subgroup fairness with respect to multiple sensitive attributes. Taking inspiration from the disentangled representation learning literature, we propose an algorithm for learning compact representations of datasets that are useful for reconstruction and prediction, but are also flexibly fair, meaning they can be easily modified at test time to achieve sub-group demographic parity with respect to multiple sensitive attributes and their conjunctions. We show empirically that the resulting encoder— which does not require the sensitive attributes for inference—enables the adaptation of a single representation to a variety of fair classification tasks with new target labels and subgroup definitions.

No abstract available.

Many selection procedures involve ordering candidates according to their qualifications.  For example, a university might order applicants according to a perceived probability of graduation within four years, and then select the top 1000 applicants.  In this work, we address the problem of ranking members of a population according to their "probability" of success, based on a training set of *binary* outcome data (e.g., graduated in four years or not).  We show how to obtain rankings that satisfy a number of desirable accuracy and fairness criteria, despite the coarseness of binary outcome data.  As the task of ranking is global -- that is, the rank of every individual depends not only on their own qualifications, but also on every other individuals' qualifications -- ranking is more subtle (and vulnerable to manipulation) than standard prediction tasks.

We develop two hierarchies of definitions of "evidence-based" rankings.  The first hierarchy relies on a semantic notion of *domination-compatibility*:  if the training data suggest that members of a set A are on-average more qualified than the members of B, then a ranking that favors B over A (i.e., where B *dominates* A) is blatantly inconsistent with the data, and likely to be discriminatory.  The definition asks for domination-compatibility, not just for a pair of sets, but rather for every pair of sets from a rich collection C of subpopulations.  The second hierarchy aims at precluding even more general forms of discrimination; this notion of *evidence-consistency* requires that the ranking must be justified on the basis of consistency with the expectations for every set in the collection C.  Somewhat surprisingly, while evidence-consistency is a strictly stronger notion than domination-compatibility when the collection C is predefined, the two notions are equivalent when the collection C may depend on the ranking in question and are closely related to the notion of *multi-calibration* for predictors.

Motivated by the need to audit complex and black box models, there has been extensive research on quantifying how data features influence model predictions. Feature influence can be direct (a direct influence on model outcomes) and indirect (model outcomes are influenced via proxy features). Feature influence can also be expressed in aggregate over the training or test data or locally with respect to a single point. Current research has typically focused on one of each of these dimensions. In this paper, we develop disentangled influence audits, a procedure to audit the indirect influence of features. Specifically, we show that disentangled representations provide a mechanism to identify proxy features in the dataset, while allowing an explicit computation of feature influence on either individual outcomes or aggregate-level outcomes. We show through both theory and experiments that disentangled influence audits can both detect proxy features and show, for each individual or in aggregate, which of these proxy features affects the classifier being audited the most. In this respect, our method is more powerful than existing methods for ascertaining feature influence.

Consequential decision-making typically incentivizes individuals to behave strategically, tailoring their behavior to the specifics of the decision rule. A long line of work has therefore sought to counteract strategic behavior by designing more conservative decision boundaries in an effort to increase robustness to the effects of strategic covariate shift. We show that these efforts benefit the institutional decision maker at the expense of the individuals being classified. Introducing a notion of social burden, we prove that any increase in institutional utility necessarily leads to a corresponding increase in social burden. Moreover, we show that the negative externalities of strategic classification can disproportionately harm disadvantaged groups in the population. Our results highlight that strategy-robustness must be weighed against considerations of social welfare and fairness. This is joint work with Smitha Milli, Anca Dragan, and Moritz Hardt.

Many decisions that once were made by humans are now made using algorithms. These algorithms are typically optimized for a single, private objective: Loan decisions are optimized for profit, smart phone apps are optimized for engagement, and news feeds are optimized for clicks. However, these decisions have side effects: irresponsible payday loans, addictive apps, and fake news may maximize private objectives, but have unclear impacts on social welfare. This project develops an approach to welfare-sensitive machine learning, which provides a formal framework for weighing the social welfare impact of an algorithm alongside traditional, private optimization criteria. We describe and evaluate different strategies for measuring and predicting these welfare impacts. In partnership with a large financial institution in Kenya, we stress-test this approach with real lending data, and characterize the empirical tradeoff between private (profit) and public (welfare) objectives in a high-stakes environment.

Since many critical decisions impacting human lives are increasingly being made by algorithms, it is important to ensure that the treatment of individuals under such algorithms is demonstrably fair under reasonable notions of fairness. One compelling notion proposed in the literature is that of individual fairness (IF), which advocates that similar individuals should be treated similarly (Dwork et al. 2012). Originally proposed for offline decisions, this notion does not, however, account for temporal considerations relevant for online decision-making. In this paper, we extend the notion of IF to account for the time at which a decision is made, in settings where there exists a notion of conduciveness of decisions as perceived by the affected individuals. We introduce two definitions: (i) fairness-across-time (FT) and (ii) fairness-in-hindsight (FH). FT is the simplest temporal extension of IF where treatment of individuals is required to be individually fair relative to the past as well as future, while in FH, we require a one-sided notion of individual fairness that is defined relative to only the past decisions. We show that these two definitions can have drastically different implications in the setting where the principal needs to learn the utility model. Linear regret relative to optimal individually fair decisions is inevitable under FT for non-trivial examples. On the other hand, we design a new algorithm: Cautious Fair Exploration (CaFE), which satisfies FH and achieves sub-linear regret guarantees for a broad range of settings. We characterize lower bounds showing that these guarantees are order-optimal in the worst case. FH can thus be embedded as a primary safeguard against unfair discrimination in algorithmic deployments, without hindering the ability to take good decisions in the long-run.

We study an online classification problem with partial feedback in which individuals arrive one at a time from a fixed but unknown distribution, and must be classified as positive or negative. Our algorithm only observes the true label of an individual if they are given a positive classification. This setting captures many classification problems for which fairness is a concern: for example, in criminal recidivism prediction, recidivism is only observed if the inmate is released; in lending applications, loan repayment is only observed if the loan is granted. We require that our algorithms satisfy common statistical fairness constraints (such as equalizing false positive or negative rates --- introduced as "equal opportunity" in Hardt et al. (2016)) at every round, with respect to the underlying distribution. We give upper and lower bounds characterizing the cost of this constraint in terms of the regret rate (and show that it is mild), and give an oracle efficient algorithm that achieves the upper bound.

Joint work with Katrina Ligett, Aaron Roth, Bo Waggoner and Steven Wu.