Analyzing and Enhancing Algorithmic Fairness in Social Systems and Data-Restricted Applications

Abstract

Over the past decade, machine learning (ML) and artificial intelligence (AI) have become increasingly prevalent in applications that impact our daily lives. However, their use in high-stakes domains involving sensitive data raises significant ethical and legal concerns, particularly around algorithmic bias. Research on fairness in AI and ML (FairAI) seeks to address how the decisions made by such models may conflict with societal values. This dissertation contributes to this effort by addressing key challenges in improving algorithmic fairness within social systems and data-constrained applications, aiming to ensure ethical model deployment in high-stakes situations. Across its three parts, this dissertation begins with algorithmic development and analysis grounded in static fairness assumptions, and later revisits the limits of those assumptions, culminating in a framework that models fairness as a dynamic sequential process shaped by temporal interventions. The first part of this dissertation proposes an algorithm to achieve both inter-group and within-group fairness in static decision-making problems. While many studies focus on fairness across different demographic groups, algorithms designed for inter-group fairness can unintentionally treat individuals within the same group unfairly. To address this issue, we introduce the notion of within-group fairness and present a pre-processing framework that satisfies both inter- and within-group fairness with minimal loss in ensemble prediction accuracy. This framework maps feature vectors from different groups to a fair canonical domain before passing them through a scoring function, preserving the relative relationships among scores within the same demographic group to guarantee within-group fairness. The second part of this dissertation explores trade-offs in satisfying multiple fairness constraints in static data-restricted decision-making contexts. While previous research has explored trade-offs between fairness and ensemble prediction accuracy through analyzing model outputs, these studies do not consider how data restrictions impact a model's ability to satisfy fairness constraints. To fill this gap, we propose a framework that models fairness-accuracy trade-offs in data-restricted settings. Our framework analyzes the optimal Bayesian classifier’s behavior using a discrete approximation of the data distribution, allowing us to isolate the effects of fairness constraints. Our results demonstrate that this framework provides an effective, structured approach for practitioners to assess fairness constraints in decision-making pipelines. Building on these insights, the third part of this dissertation shifts its focus to analyzing fairness from a sustainability perspective. Prior research has shown that applying fairness constraints to static, single-stage decision-making problem formulations can have negative long-term effects on disadvantaged groups. Recognizing that fairness interventions unfold over time and often involve multiple decision points rather than isolated decisions, we develop the notion of Multi-Agent Fair Environments (MAFEs)—testbeds for evaluating FairAI algorithms in temporally evolving social systems. We then present and analyze three MAFEs that model distinct social systems. We model each decision point as an agent within these MAFEs, leveraging their dynamic interactions to enable greater flexibility and more insightful analysis of system dynamics. Experimental results demonstrate the utility of our MAFEs as testbeds for developing multi-agent fair algorithms.