Speaker: Nur Kaynar (Assistant professor of Operations, Technology, and Information Management at the SC Johnson Graduate School of Management at Cornell University)
Title: Causal Product Networks for Basket Shopping: Discovery and Applications
Date: 20th of September , 2024 (Friday)
Time: 13:30 – 14:30
Place: EA409
Understanding how the purchase of one product affects the purchase of others in shopping baskets is crucial for retailers to effectively tailor their assortment strategies. However, discovering and modeling these interactions is challenging due to the exponentially large number of possible interactions among numerous products across different categories. Our study addresses this challenge by using a causal discovery approach that learns the underlying causal structure among product purchases from observational shopping basket data, filters out non-causal correlations, and constructs a causal product network to describe these latent interactions. We validate this approach by empirically investigating the behavior of basket-shopping consumers utilizing a large-scale basket shopping dataset. Our study provides three main empirical findings. First, we investigate how the proposed causal model performs in describing the relationships among product purchases compared to other network structures that represent hypothetical consumer behavior. Our results suggest causal product networks represent product interactions in a shopping basket more accurately compared to models using all possible connections and based on correlations among product purchases, thereby reducing complexity.
Second, we investigate various specifications of product interactions used in previous literature, including both product-level and category-level interactions, and find that considering product level interactions most accurately capture the dynamics of basket shopping.
Third, we compare the disparity between brick-and-mortar and online channels regarding the sparsity of causal relationships and discover that the online channel exhibits fewer causal connections among product purchases. Using the constructed causal product networks, we demonstrate their application in an assortment optimization problem in the context of basket shopping. Our model outperforms classical Multinomial Logit (MNL) models by around 20%-43% in total sales. These results indicate that the proposed causal product network model can accurately describe the purchase behavior of basket-shopping customers with fewer parameters and aid in developing effective assortment strategies.
Bio:
Nur Kaynar is an assistant professor of Operations, Technology, and Information Management at the SC Johnson Graduate School of Management at Cornell University. Her research is at the intersection of optimization, causal inference, and operations. Specifically, she is interested in blending techniques from econometrics and machine learning to achieve efficient causal inference methods. At Johnson, she has taught elective courses on programming with Python and SQL in the full-time MBA program. Before joining Cornell, she received a Ph.D. from the Anderson School of Management at UCLA and completed both her master’s and undergraduate degrees at Bilkent University.