Understanding and Mitigating Bias in Algorithms, Data, and Society
Author(s)
Zhao, Zhanzhan
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Societal biases are systemic prejudices embedded in individuals’ and institutions’ collective beliefs, attitudes, norms, and behaviors, including racial biases, ideological biases, and partisan biases. While computer scientists have made significant strides in quantifying patterns of bias and developing mitigation strategies for some computer algorithms and internet applications, exploring the root causes of most societal biases remains a challenge. Due to the complex nature and unpredictability of emergent societal behaviors and the limited data availability, many powerful algorithms and data analysis tools face inherent limitations.
This thesis develops new algorithms and data analytical tools for identifying and addressing societal biases. We take a multi-lens approach, combining algorithmic modeling, data analysis, and qualitative theories to gain more comprehensive insights into societal biases along three specific themes: the persistent residential segregation arising from racial homophily, ideological polarization exacerbated by reinforcements of beliefs, and non-responsiveness in political elections boosted by partisan redistricting policies. By integrating these disparate research methods, we aim to develop a deeper understanding of these equity-related social problems and propose system-level mitigation interventions.
First, using tools from probability and theoretical computer science, we rigorously examine the possible causal mechanisms behind residential segregation through Markov chain analysis. We prove how the careful placement of incentives, such as urban amenities, can either worsen or alleviate segregation, supported by machine learning analysis of U.S. cities. Second, we explore how personalized recommender algorithms can contribute to extreme polarization through agent-based modeling. Our simulations highlight the effectiveness of novel local recommendation designs in reversing polarization, even in the presence of extreme influencers and social pressure. We emphasize the importance of tailored interventions that leverage diversity to foster inclusive dialogue and bridge ideological divides. Finally, we also develop a fair evaluation method for biases in public policies, proposing a novel computational protocol with improved efficiency. Our study of Georgia's districting plans uncovers their inadequate responsiveness to changing voter opinions, emphasizing the instrumental role of algorithms and data analysis in auditing and bringing clarity to public policies.
Sponsor
Date
2023-07-30
Extent
Resource Type
Text
Resource Subtype
Dissertation