Degree
Doctor of Philosophy (PhD)
Department
Computer Science and Engineering
Document Type
Dissertation
Abstract
As Artificial Intelligence (AI) increasingly governs resource allocation, including item exposure in ranking, public services in geographic regions, and computational workloads in distributed systems, ensuring algorithmic equity has become a primary technical challenge. This thesis addresses the limitation of ``one-size-fits-all'' fairness metrics by arguing that equity must be mathematically reformulated to align with the unique structural constraints of a given domain.
We investigate this hypothesis across three high-impact areas:
- Information Retrieval: We address the disparity caused by Top-K truncation. By introducing a novel Top-K Exposure Disparity measure, we transform the non-differentiable selection process into a tractable objective that ensures fairness at the point of decision.
- Spatio-temporal Forecasting:}We tackle the ``boundary problem'' where global metrics obscure local inequities. Our framework, LG-Fair, optimizes a multi-scale hybrid loss function to reduce bias along socioeconomic boundaries in mobility and public safety models.
- Distributed Systems: We extend equity to Load Balancing in Federated Mixture-of-Experts (MoE). We develop FLEX-MoE, an adaptive algorithm that optimizes expert utilization under hardware constraints to ensure system efficiency through equitable load distribution.
Our results demonstrate that tailoring fairness constraints to domain-specific requirements, such as top-K ranked list truncation, neighborhood boundaries, or hardware constraints, achieves a superior balance between system performance and prediction fairness.
Date
8-30-2026
Recommended Citation
Zhang, Boyang, "Domain-Specific Fairness-Aware AI in Resource Allocation Systems" (2026). LSU Doctoral Dissertations. 7196.
https://repository.lsu.edu/gradschool_dissertations/7196
Committee Chair
Sun Mingxuan
LSU Acknowledgement
1
LSU Accessibility Acknowledgment
1