Towards Trustworthy Machine Learning: Fairness Under Dynamics and Counterfactuals

Machine learning systems increasingly make decisions with long-term consequences for individuals and society. Designing trustworthy machine learning systems therefore requires going beyond static accuracy and one-shot fairness criteria. Decisions may alter future populations through feedback, individuals may improve and reapply, uncertainty may call for set-valued predictions rather than point estimates, and large-scale deployment may be constrained by the efficiency of modern generative models. This thesis explores these challenges through contributions spanning fairness under dynamics, counterfactual fairness under predictive uncertainty, and efficient inference serving for large language models. We first study fairness in dynamic decision-making settings such as lending, hiring, and admissions, where rejected individuals may invest effort to improve and reapply. We introduce Equal Improvability, a fairness notion requiring that rejected individuals from different groups have comparable opportunities to become accepted after bounded effort. We develop optimization-based methods for learning classifiers that satisfy this criterion and show that they can promote more equitable long-term population dynamics. Next, we consider sequential selection by multiple decision-makers from a shared applicant pool. We propose fair-greedy policies that balance utility with long-term representation targets, characterize convergence under positive reinforcement dynamics, and show that policy design must account for more complex feedback mechanisms to avoid unintended negative reinforcement. We then address fairness under counterfactual reasoning and predictive uncertainty. Counterfactually Fair Conformal Prediction constructs prediction sets that are both statistically valid and counterfactually fair. By aggregating conformity scores across counterfactual interventions on protected attributes, the method produces prediction sets that remain invariant under counterfactual changes to the protected attribute while preserving finite-sample marginal coverage guarantees. Finally, we study computational efficiency in generative machine learning systems. Multi-Bin Batching is an inference-serving policy for large language models that groups requests by predicted generation lengths to reduce inefficiency from heterogeneous runtimes. Queueing-theoretic analysis characterizes the throughput benefits of length-aware batching and shows that the effective service rate approaches the idealized static-batching capacity as the number of bins increases. Experiments on simulated and real inference workloads demonstrate substantial throughput improvements over static batching. Together, these contributions provide fairness notions and algorithms that remain meaningful under feedback, uncertainty, and counterfactual interventions, while also developing principled serving mechanisms that support scalable deployment of generative models.