Trustworthy decision-making in data-driven optimization: Fairness, generalization, and interpretability for revenue management and inventory control in e-commerce

Xu, Zexing

Permalink

https://hdl.handle.net/2142/129463 Copy

Description

Title

Trustworthy decision-making in data-driven optimization: Fairness, generalization, and interpretability for revenue management and inventory control in e-commerce

Author(s)

Xu, Zexing

Issue Date

2025-05-02

Director of Research (if dissertation) or Advisor (if thesis)

Etesami, Rasoul

Doctoral Committee Chair(s)

Etesami, Rasoul

Committee Member(s)

• Tong, Hanghang
• Wang, Qiong
• Marla, Lavanya
• Zhang, Linjun

Department of Study

Industrial&Enterprise Sys Eng

Discipline

Systems & Entrepreneurial Engr

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Data-Driven Optimization
• Trustworthy AI
• Decision Making
• Fairness
• Generalization
• Robustness
• Interpretability
• E-Commerce

Abstract

The increasing reliance on data-driven models in e-commerce and revenue management has brought to the forefront the critical need for trustworthy optimization techniques that not only maximize performance but also ensure fairness, robustness, and interpretability. This dissertation addresses these needs by introducing the first unified optimization framework that embeds key trustworthiness criteria and exploring advanced data-driven optimization methodologies, focusing on fairness, out-of-domain generalizability, and interpretability. The first part of the thesis delves into the development of optimization frameworks that incorporate fairness constraints, particularly in the context of personalized pricing within e-commerce. By leveraging multiple distance metrics, we establish group fairness constraints designed to mitigate potential biases in pricing models. These fairness constraints are then integrated into a stochastic optimization model aimed at maximizing revenue, which we simplify to a linear program for efficient computation. Our theoretical analysis reveals the challenges in balancing fairness with revenue optimization, particularly in scenarios where multiple fairness constraints conflict. We explore the implications of these constraints on revenue, social welfare, and consumer surplus, offering a comprehensive analysis of how fairness-oriented pricing strategies can influence key business outcomes. The second part investigates the challenge of generalization, specifically focusing on Out-of-Domain (OOD) generalization in dynamic e-commerce environments. To address the limited availability of historical data during peak periods, we propose a meta-learning approach that leverages proxy data from non-peak periods, enriched by features learned from Graph Neural Networks (GNNs). This method enhances the robustness of demand forecasting models, allowing them to generalize better to OOD data during high-stakes sales events. Our theoretical analysis demonstrates that by considering domain similarities through task-specific metadata, the model achieves improved generalization, with excess risk decreasing as the number of training tasks increases. Empirical evaluations on large-scale industrial datasets underscore the superiority of this approach, with significant improvements in demand prediction accuracy over state-of-the-art models. Interpretability is the focus of the third part of the thesis, addressing the need for transparent models that stakeholders can trust and understand in e-commerce and inventory management. We first embed rich contextual features in a multi-period inventory framework and devise a \emph{Contextual Value Iteration} (CVI) algorithm that both provably converges and yields human-readable policies—explicitly linking each feature vector to its optimal order decision. Empirical studies confirm CVI’s superior efficiency and robustness over vanilla MDP and cMDP baselines. We then extend interpretability from algebraic structure to \emph{executable code}. To this end, we release the textbf{FaithfulPersonaCodeX} benchmark—169 Python solutions paired with ten Stack Overflow–derived user profiles, and propose the \textbf{DISCO} pipeline, which fuses faithfulness-preserving self-critique with persona-aware rewriting. DISCO improves GPT-3.5 \textsc{Pass@5} by 3.7\% over Self-Consistency and achieves a 61.08\% win rate in LLM-as-Judge personalization tests, outperforming commercial code-explanation tools despite requiring no additional training. Finally, the practical implications of the proposed methodologies are explored through case studies in e-commerce, particularly focusing on revenue management and inventory control. These applications demonstrate the effectiveness of our approaches in real-world scenarios, underscoring the potential for trustworthy data-driven optimization to drive sustainable and ethical business practices. This dissertation contributes to the advancement of data-driven optimization by integrating fairness, robustness, and interpretability into learning algorithms, paving the way for more reliable and equitable applications in e-commerce, particularly focusing on revenue management and inventory control. Moreover, this work lays a versatile foundation for future extensions, such as jointly enforcing fairness, reliability, and transparency constraints, and has influenced production systems at Amazon, Meta, and Google, underscoring its practical impact.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129463

Copyright and License Information

Copyright 2025 Zexing Xu

Owning Collections