Preventing Disparate Treatment in Sequential Decision Making

Abstract

We study fairness in sequential decision making environments, where at each time step a learning algorithm receives data corresponding to a new individual (e.g. a new job application) and must make an irrevocable decision about him/her (e.g. whether to hire the applicant) based on observations made so far. In order to prevent cases of disparate treatment, our time-dependent notion of fairness requires algorithmic decisions to be consistent: if two individuals are similar in the feature space and arrive during the same time epoch, the algorithm must assign them to similar outcomes. We propose a general framework for post-processing predictions made by a black-box learning model, that guarantees the resulting sequence of outcomes is consistent. We show theoretically that imposing consistency will not significantly slow down learning. Our experiments on two real-world data sets illustrate and confirm this finding in practice.

Individual Fairness in Hindsight

“…Only recently has the literature started looking at the implications of fairness constraints in online learning settings. Our model and results complement a small but growing body of literature in this domain (Joseph et al 2016b, Liu et al 2017, Gillen et al 2018, Heidari and Krause 2018, Celis et al 2018, Joseph et al 2016a, Elzayn et al 2018). The two papers most related to our work are Joseph et al (2016b) and Heidari and Krause (2018) that we discuss below.…”

“…Our model and results complement a small but growing body of literature in this domain (Joseph et al 2016b, Liu et al 2017, Gillen et al 2018, Heidari and Krause 2018, Celis et al 2018, Joseph et al 2016a, Elzayn et al 2018). The two papers most related to our work are Joseph et al (2016b) and Heidari and Krause (2018) that we discuss below. Joseph et al (2016b) were one of the earliest to study the impact of fairness constraints on learning in a contextual multi-armed bandit setting under a utility maximization objective.…”

“…In this paper, we consider the notion of individual fairness (Dwork et al 2012), that relies on the premise of equality of treatment, requiring that "similar" individuals must be treated "similarly". This intuitively compelling notion of fairness was proposed in the influential work of Dwork et al (2012) in the context of classification in supervised learning, and has since been studied under several settings (see Yona and Rothblum 2018, Dwork and Ilvento 2018, Heidari and Krause 2018. The key idea is to introduce a Lipschitz condition on the decisions of a classifier, such that for any two individuals x, y that are at distance d(x, y), the corresponding distributions over decisions M (x) and M (y) are also statistically close within a distance of some multiple of d(x, y).…”

Designing fair AI for managing employees in organizations: a review, critique, and design agenda

“…In the organizational justice literature, distributive justice is often assessed either by comparing an individual's inputs to their outputs (i.e., within individuals) or by comparing an individual's outcomes to others' outcomes (i.e., between individuals or groups). In our review of the AI fairness literature, we found that distributive fairness was overwhelmingly measured by comparing outcomes across people rather within the individual (e.g., Glymour & Herington, 2018 � ;Grgić-Hlača et al, 2018;Heidari & Krause, 2018). For example, Heidari and Krause (2018) argued that fairness should be measured by determining whether similar individuals should be assigned similar outcomes by the algorithm.…”

“…In our review of the AI fairness literature, we found that distributive fairness was overwhelmingly measured by comparing outcomes across people rather within the individual (e.g., Glymour & Herington, 2018 � ;Grgić-Hlača et al, 2018;Heidari & Krause, 2018). For example, Heidari and Krause (2018) argued that fairness should be measured by determining whether similar individuals should be assigned similar outcomes by the algorithm. This focus on assessing fairness among individuals or groups might be from the AI fairness literature's tendency to measure distributed fairness from the outside (i.e., detecting disparate impact of outcomes via a mathematically validated algorithm or legal standard) rather than from the perspective of the impacted individual.…”