The Role of Noise, Proxies, and Dynamics in Algorithmic Fairness

Machine learning is increasingly used to aid or automate decision making. Yet, algorithmic solutions often suffer from bias and disparate impact across demographic groups. For many application settings, the mechanisms by which bias arises and the effects of applying fairness-aware learning methods are not sufficiently understood. In this thesis, I focus on differential noise and missingness as drivers of bias (Part I), and long-term dynamics of fairness promoting interventions (Part II). 

Part I of the thesis presents three studies on the impacts of differentially missing observations, differential feature mismeasurement, and differentially informative proxies. First, we discuss how geographical differences in victim crime reporting rates can lead to outcome disparities in predictive policing systems. Second, we explore the fairness implications of differential feature under-reporting in the setting of public sector risk assessment instruments, and propose technical solution approaches. The third study proposes a sandbox tool to evaluate fairness-enhancing algorithms under different types of artificially injected bias. Potential use cases for the tool are demonstrated via case studies. 

Part II of the thesis comprises a study of long-term dynamics of fairness intervention in connection recommendation. Using both simulation and theoretical limit analysis, we demonstrate how typical fairness-promoting interventions can fail to promote equity in second order variables of interest such as network sizes.