CEE-CSE PhD Thesis Defense: Junyi Sha

Abstract:

Artificial intelligence (AI) is increasingly central across industries, yet its application to operations management and decision-making extends beyond classical tabular forecasting and introduces new challenges. Three recurring gaps arise: adapting general purpose models to domain-specific settings with limited and unstructured data; ensuring interpretability so outputs can inform decisions; and avoiding the loss of diversity when optimizing on historical outcomes. This thesis examines these challenges in fashion retail, focusing on how AI can support both accurate prediction and interpretable, actionable design decisions. In parallel, simple game environments are used to study the trade-off between optimality and diversity in AI-driven decision-making.

Traditional demand forecasting models rely primarily on structured numerical and categorical features and often overlook the importance of visual or textual information due to the lack of effective methods to extract and use them. Meanwhile, fast fashion retailers must continuously introduce new products with no sales history and extremely short life cycles, which leaves little opportunity for merchants to test new designs before launch or to replenish existing ones once demand is observed. We investigate these settings in Chapters 2-3. In Chapter 2, we compare different existing machine and deep learning model architectures, and introduce effective hybrid model approaches that significantly improve demand forecasting accuracy. Specifically, we borrow the idea of image similarity from recommendation systems and incorporate similarity derived features into a gradient-boosting model. Chapter 3 extends the multimodal forecasting model from a point prediction to a ranking system that can take in different design information and output the relative popularity of the given designs. Using choice modeling, we estimate both the marginal effect of each individual design feature and the interaction effects between features. These insights are then used to guide the design of new products and the refinement of existing ones through diffusion models. Furthermore, we show the potential loss of diversity in AI-driven decision spaces. In Chapter 4, we study the trade-off between optimality and diversity using large language models in simple game environments. We illustrate that without careful design, the learning process may easily collapse into optimizing for winrate at the expense of diversity. With careful data preparation, explicitly encouraging diversity during training can further enhance model performance in our settings.

In summary, this thesis contributes multimodal forecasting and design methods together with complementary evidence on optimality versus diversity, advancing more interpretable, actionable, and diversity-aware uses of AI in operations management.

Thesis Advisor:

• David Simchi-Levi, William Barton Rogers Professor in Energy, Professor of Engineering Systems