A Feature-Based Consideration Set Choice Model for Online Retailing

Problem definition : We propose a feature-based consideration set choice model (FCM) motivated by customers’ purchasing behavior in online retail platforms. In our model, customers form a consideration set by including products with the highest utilities computed based on a subset of product features visible on the search results page. The customers then make a purchase decision from the consideration set by accounting for all features available on the product pages. The FCM incorporates heterogeneity in customer preferences across both the consideration and purchase stages, and it allows customers to re-evaluate products’ utility in the second stage. Methodology/results : We develop an efficient maximum likelihood estimation procedure for estimating the model parameters from customers’ click and purchase data. We show that the assortment optimization problem under FCM is NP-hard by drawing a connection to the assortment problem under the latent-class multinomial logit (LC-MNL) model. Moreover, we establish that the problem remains NP-hard even under a variant of our model termed the deterministic feature-based consideration set choice model (D-FCM), which imposes a deterministic structure in the consideration stage. On the positive side, the D-FCM admits a tractable mixed integer linear programming (MILP) formulation for solving the assortment problem and facilitates the study of the more complex joint assortment and pricing problem, for which we provide a polynomial-time solution (in the number of products) for both homogeneous and heterogeneous settings. Managerial implications : Through numerical experiments on real and synthetic data, we demonstrate that our proposed model outperforms the MNL and LC-MNL benchmarks on out-of-sample prediction accuracy and decision performance. Finally, we introduce a novel operational problem termed feature selection , which identifies the subset of features to display during the consideration formation stage to maximize expected revenue. We establish the NP-hardness of this problem under both model variants and propose a tractable MILP formulation for solving it.