Maggie Wu's Dissertation Defense
PhD candidate Maggie Wu will present her dissertation, "Harmonizing Fairness and Utility in Machine Learning." Her final examination committee includes Professor Jingrui He (Chair); Associate Professor Nigel Bosch; Assistant Professor Jiaqi Ma; Jiawei Han, professor in the Siebel School of Computing and Data Science; and John Birge, professor in the Booth School of Business at the University of Chicago.
Abstract
Machine learning systems must often balance predictive utility with fairness across demographic groups. This challenge becomes more difficult when outcome labels are incomplete, sensitive attributes are sparsely observed, groups are unevenly represented, or model access limits the available interventions. This dissertation develops five complementary methods for addressing these constraints, organized around data-centric and model-centric perspectives. From the data-centric perspective, FairPUL provides a model-agnostic post-processing framework for fair positive and unlabeled learning. Under its labeling and estimation assumptions, the framework supports equal opportunity or equalized odds and establishes consistency with respect to classification risk and the selected fairness criterion. FairIF uses influence-guided sample reweighting to improve group fairness when sensitive attributes are available only in a validation set. FADIG addresses group imbalance in unsupervised anomaly detection through adaptive reconstruction weighting and fairness-aware contrastive learning. From the model-centric perspective, FairGM aligns group-conditioned fairness information at the levels of metric values, mean gradients, and gradient variances. A preference-aware multi-objective formulation makes the balance between these objectives and predictive utility explicit. FairIF also addresses architectural constraints by preserving the target architecture while changing training-sample influence. FairRec extends the investigation to LLM recommendation through item-level decoding control. It combines a fixed calibrated disparity threshold with hard constraints or an adaptive soft penalty, without retraining the underlying model. The dissertation provides theoretical analyses under method-specific assumptions and empirical evaluations across various tasks. The reported results show that interventions tailored to the available information and source of disparity can improve measured fairness while retaining competitive task performance. These contributions clarify both the opportunities and the limits of harmonizing fairness and utility under practical data and model constraints.
Questions? Contact Maggie Wu.