Dynamics of Algorithmic Fairness
CitationHu, Lily. 2022. Dynamics of Algorithmic Fairness. Doctoral dissertation, Harvard University Graduate School of Arts and Sciences.
AbstractThe rise of machine learning-based predictive models in making decisions of profound social impact has spurred study of those technical properties that may bear on the moral and political character of their deployment. One area of normative concern that has received particularly heightened scrutiny is the fairness of the outcomes that data-based classification tools issue. Much computer science and mathematics-based research in the fields of algorithmic fairness and fair machine learning looks to diagnose when classifiers may be engaging in discrimination or otherwise generating unfair outputs and to prevent such outcomes using a variety of methods that seeks to alter the classifier's behavior and thus the outcomes it produces.
This dissertation comprises contributions to the burgeoning field of algorithmic fairness that, rather than focusing on the internal workings of an algorithmic system itself, centers instead the interaction between machine classifications and the broader societal contexts within which data-based predictive tools are embedded. Each of these works thus conceive of algorithmic tools as only one component of a larger sociotechnical system that distributes key social benefits and burdens. Over the span of the three projects contained within—"Disparate Effects of Strategic Manipulation," "A Short-term Intervention for Long-term Fairness," and "Fair Classification and Social Welfare"—it considers changes to institutional incentive structures that data-based classification introduces, the strategic responses of agents who interact with such systems, and the welfare impacts of various fairness constraints that have been proffered in the field. Approaching the fairness problem with this wider lens of analysis builds in a broader and longer-term perspective from the start and necessarily draws on methods and insights beyond that of applied mathematics and computer science. In so doing, this research makes distinctive contributions to matters that are central in the scholarly discourse in algorithmic fairness, such as debate about fairness-accuracy trade-offs in algorithmic decision-making and the strategic interplay between machine classifications and agent behaviors. This dissertation therefore both advocates for and itself exemplifies a reorientation to questions of fairness by shifting focus from the machine as the central object of interest in favor of a broader vantage that addresses the broader social dynamics of algorithmic fairness. This approach not only challenges the standard methodological tacks taken in the field of algorithmic fairness but also generates insights that track more closely to how these tools actually operate in the world to effect key social outcomes. It thus is better suited to guiding work in algorithmic fairness towards the kinds of interventions we will need to construct a more equitable society.
Citable link to this pagehttps://nrs.harvard.edu/URN-3:HUL.INSTREPOS:37372041
- FAS Theses and Dissertations