Date of Graduation

8-2025

Document Type

Thesis

Degree Name

Master of Science in Statistics and Analytics (MS)

Degree Level

Graduate

Department

Statistics and Analytics

Advisor/Mentor

Zhang , Lu

Committee Member

Chakraborty, Avishek

Second Committee Member

Wu, Xintao

Keywords

algorithmic recourse; long-term fairness

Abstract

Algorithmic decision-making systems are increasingly being deployed in high-stakes domains such as criminal justice, education, and financial services. While machine learning models have demonstrated significant utility in automating complex decisions, they have also raised substantial concerns regarding fairness and equity. Much of the existing work in algorithmic fairness has focused on static, one-shot settings, where interventions are aimed at miti- gating bias in a single decision. However, many real-world systems operate sequentially, where decisions made at one point in time can influence future outcomes through dynamic feedback loops. In such scenarios, addressing fairness at only one decision point can be insufficient or even counterproductive, as it may ignore the cumulative and compounding nature of disadvantage over time. This thesis addresses the challenge of achieving long-term fairness in sequential decision-making through a causal recourse framework. Our key contri- bution lies in offering a novel perspective that complements the literature by bridging causal inference, generative modeling, and fairness-aware optimization to develop a principled ap- proach for mitigating disparities over time. Specifically, we propose a two-part framework named SCARF (Sequential Causal Algorithmic Recourse for Fairness), which combines a Variational Causal Graph Autoencoder (VACA) and a Recurrent Conditional Generative Adversarial Network (RC-GAN) to model and simulate how interventions influence future outcomes through causal pathways. VACA captures path-specific effects within a struc- tural causal model, allowing us to identify actionable and temporally sensitive interventions. Meanwhile, RC-GAN models time-lagged dependencies in the data, enabling the genera- tion of counterfactual trajectories that evolve under specified interventions. Our problem is formulated as a constrained optimization task that balances predictive utility with fairness constraints over a finite decision horizon. Interventions are restricted by a predefined budget, reflecting the practical cost of altering an individual’s features or opportunities. The opti- mization objective incorporates both short-term fairness and long-term fairness goals over the trajectory of decisions. The framework supports counterfactual simulations that allow us to trace the downstream impact of fairness interventions, identify path-dependent inequali- ties, and evaluate intervention strategies in terms of both effectiveness and cost. Empirical evaluation is conducted on both synthetic and real-world datasets, including benchmark data in algorithmic fairness literature. Results show that the proposed method significantly reduces group disparities in long-term outcomes without sacrificing predictive performance. The contributions of this thesis are threefold. First, we introduce a novel integration of structural causal models and generative sequence models for fairness-aware decision-making. Second, we provide a fairness-aware optimization procedure that operates under realistic constraints and supports principled recourse planning. Third, we offer empirical insights into the trade-offs between fairness and intervention cost over time. By unifying these el- ements, this thesis advances the study of long-term algorithmic fairness and highlights the importance of causal reasoning in designing equitable decision-making systems that extend beyond single outcomes to entire life trajectories.

Download

Share

COinS