Learning from Consideration Sets

Département Information Systems et Operations Management

Intervenant: Canan Ulu (Georgetown)

Salle Bernard Ramanantsoa

Prior literature on consumer behavior suggests that consumers engage in a two-stage shopping behavior: in the first stage, they consider only a subset of the products offered for purchase (known as the consideration set), and in the second stage, they make a final purchase decision among the considered products based on their preferences. We model such purchase behavior using the Random Consideration Set (RCS) model (Manzini and Mariotti, 2014). In this model, consumers consider each product independently with a given consideration probability. Motivated by environments in which consumers’ consideration sets are observable (e.g., based on clickstream, eye-tracking, heatmap data sources), we consider how a decision maker should design product assortments to maximize profit while also learning about consumers’ consideration probabilities over a finite time horizon. We show that the structure of the optimal assortment depends on two orders: the consumers’ preference order and the product “informativeness” order, which we formalize using Blackwell sufficiency (Blackwell, 1951). The optimal assortment has the well-known popular set characterization when the consumers’ preference order and the product informativeness order are identical. Otherwise, the optimal assortment is popular within the set of products over which the two orders agree—a generalization of the popular set result. Based on our numerical experiments, we find that the decision maker’s profit can increase by up to 2.62% by learning, and more than 50% of that benefit can be realized by learning from the consumers’ consideration sets. The structural properties of the optimal assortment also reduce the search space for the optimal solution, leading to a reduction of up to 97.52% in the computational time compared to a total enumeration benchmark.