Abstract
In the era of data-driven decision-making, feature-based inventory analysis has emerged as a powerful approach for enhancing operational efficiency by leveraging contextual data. However, the growing concerns over data privacy—especially with sensitive customer information—pose significant challenges to the practical deployment of such methods. This dissertation addresses this critical gap by developing a comprehensive framework for privacy-preserving inventory control under differential privacy (DP) constraints.
The dissertation comprises two main projects and one minor project. The first project focuses on the offline setting, where a differentially private algorithm is developed to solve the classical newsvendor problem. The proposed method utilizes smoothed empirical risk minimization and clipped noisy gradient descent within the Gaussian differential privacy (GDP) framework. The algorithm not only guarantees rigorous privacy protection but also achieves favorable statistical performance with provably low excess risk. The second project investigates the online setting, where data arrive sequentially and decisions must be made in real time. A novel differentially private stochastic sub-gradient method is proposed to address two fundamental challenges: censored demand and non-perishable inventory. The algorithm is designed to satisfy a strengthened form of local differential privacy (LDP). Despite the complexity introduced by data dependence and partial observability, the algorithm attains a sublinear regret bound. Overall, this work contributes to the fields of operations management and privacy-preserving machine learning by bridging theoretical rigor and practical relevance. It demonstrates that effective and privacy-compliant inventory control is achievable in both static and dynamic environments, paving the way for secure data utilization in modern supply chains. My dissertation also includes a minor project related to policy evaluation in reinforcement learning (RL).