Primary Submission Category: Design of Experiments
Dynamic Assortment Optimization via Optimal Experimental Design
Authors: Aurelien Bibaut, Guy Aridor, Nathan Kallus,
Presenting Author: Aurelien Bibaut*
Online streaming platforms hope to offer catalogs of items (music, video, etc.) that are valuable to users. We consider evaluating catalogs in terms of user choice among the included items and an outside option and study the problem of designing batched adaptive experiments that would be maximally informative about a parameter of interest in terms of this unknown choice model, such as the value of a counterfactual catalog. Our discrete choice model is parametrized by a high-dimensional matrix of utility parameters with an underlying low rank structure, capturing heterogeneity in preferences with manageable statistical complexity without a priori clusterings. We derive the semiparametric efficiency bound for parameter of interest given data from different experimental designs, yielding an interpretable criterion for selecting a design. A key technical challenge arises because the optimal experimental design depends on unknown nuisance parameters, which must be learned adaptively. We show that our approach achieves, up to a negligible term, the same regret as an oracle that knows these nuisance parameters in advance. We demonstrate the value of our approach in a real-world application to a large streaming company.