Data-Driven Learning Models with Applications to Retail Operations

Thumbnail Image




Bernstein, Fernando

Journal Title

Journal ISSN

Volume Title

Repository Usage Stats



Data-driven approaches to decision-making under uncertainty is at the center of many operational problems. These are problems in which there is an element of uncertainty (e.g., customer demand) that needs to be estimated (learned) from data (e.g., customer transaction data) in order to make online (dynamic) operational (e.g., assortment) decisions. This dissertation adopts a data-driven active learning approach to study various operational problems under uncertainty with a focus on retail operations.

The first two essays in this dissertation study the classic exploration (i.e., parameter estimation) versus exploitation (i.e., optimization) trade-off from different perspectives. The first essay takes a theoretical approach and studies such trade-off in a combinatorial optimization setting. We show that resolving the exploration versus exploitation trade-off efficiently is related to solving a Lower Bound Problem (LBP), which simultaneously answers the questions of what to explore and how to do so. We establish a fundamental limit on the asymptotic performance of any admissible policy that is proportional to the optimal objective value of the LBP problem. We also propose near-optimal policies that are implementable in real-time. We test the performance of the proposed policies through extensive numerical experiments and show that they significantly outperform the relevant benchmark.

The second essay considers the dynamic assortment personalization problem of an online retailer facing heterogeneous customers with unknown product preferences. We propose a prescriptive approach, called the dynamic clustering policy, for dynamic estimation of customer preferences and optimization of personalized assortments. We test the proposed approach with a case study based on a dataset from a large Chilean retailer. The case study suggests that the benefits of the dynamic clustering policy under the Multinomial Logit (MNL) model can be substantial and result (on average) in more than 37% additional transactions compared to a data-intensive policy that treats customers independently and in more than 27% additional transactions compared to a linear-utility policy that assumes that product mean utilities are linear functions of available customer attributes.

Further focusing on retail operations, the final essay studies the interplay between a retailer's return and pricing policies and customers' purchasing decisions. We characterize the retailer's optimal prices in the cases with and without product returns and determine conditions under which offering the return option to customers increases the retailer's revenue. We also explore the impact of accounting for product returns on demand estimation. The preliminary numerical results based on a real dataset suggest that our model, which accounts for product returns, increases demand estimation accuracy compared to models that do not consider product returns in their estimation.





Modaresi, Sajad (2018). Data-Driven Learning Models with Applications to Retail Operations. Dissertation, Duke University. Retrieved from


Dukes student scholarship is made available to the public using a Creative Commons Attribution / Non-commercial / No derivative (CC-BY-NC-ND) license.