On the Long-term Impact of Algorithmic Decision Policies: Effort Unfairness and Feature Segregation through Social Learning

Abstract

Most existing notions of algorithmic fairness are one-shot: they ensure some form of allocative equality at the time of decision making, but do not account for the adverse impact of the algorithmic decisions today on the long-term welfare and prosperity of certain segments of the population. We take a broader perspective on algorithmic fairness. We propose an effort-based measure of fairness and present a data-driven framework for characterizing the long-term impact of algorithmic policies on reshaping the underlying population. Motivated by the psychological literature on social learning and the economic literature on equality of opportunity, we propose a micro-scale model of how individuals may respond to decision making algorithms. We employ existing measures of segregation from sociology and economics to quantify the resulting macroscale population-level change. Importantly, we observe that different models may shift the groupconditional distribution of qualifications in different directions. Our findings raise a number of important questions regarding the formalization of fairness for decision-making models.

Fair Machine Learning Under Partial Compliance

“…Our work contributes to similar e orts in fair machine learning literature towards broadening the scope of analysis to include these e ects [38,28,17]. Moreover, in exploring the impact of these dynamics, our work goes beyond assessments of algorithmic performance in static settings, furthering research on the long-term impact of proposed interventions [30,29,37].…”

Stateful Strategic Regression

“…Liu et al [32] and Kannan et al [26] study how a utility-maximizing decision-maker may respond to the predictions made by a predictive rule (e.g., the decision-maker may interpret/utilize the predictions a certain way or decide to update the model entirely.) Mouzannar et al [35] and Heidari et al [21] propose several dynamics for how individuals within a population may react to predictive rules by changing their qualifications. Dong et al [14], Hu et al [24], Milli et al [34] address strategic classification-a setting in which decision subjects are assumed to respond strategically and potentially untruthfully to the choice of the predictive model, and the goal is to design classifiers that are robust to strategic manipulation.…”

Bandit Learning with Delayed Impact of Actions

“…For algorithmic fairness sequential setting, one line of works have studied fairness in the sequential learning setting while not considering long-term impact of actions Joseph et al [2016], Bechavod et al [2019], Liu et al [2017], Gupta and Kamble [2019], Gillen et al [2018]. For the study on delayed impacts of actions, recent works mostly focus on addressing the one-step delayed impacts or a multi-step sequential setting with full information [Heidari et al, 2019, Hu and Chen, 2018, Liu et al, 2018, Mouzannar et al, 2019, Cowgill and Tucker, 2019, Bartlett et al, 2018. Our work differs from the above and studies delayed impacts of actions in sequential decision making under uncertainty.…”

“…While being a relatively under-explored (but important) topic, several recent works have looked into this problem of delayed impact of actions in algorithmic fairness, and the results have focused on understanding the impact in a one-step delay of actions [Liu et al, 2018, Kannan et al, 2019, Heidari et al, 2019, or a sequential decision making setting without uncertainty [Mouzannar et al, 2019, Zhang et al, 2019, Hu and Chen, 2018.…”