Assortment Optimization with Multi-Item Basket Purchase under the Multivariate MNL Model

Abstract

Problem definition: Assortment selection is one of the most important decisions faced by retailers. Most existing papers in the literature assume that customers select at most one item out of the offered assortment. While this is valid in some cases, it contradicts practical observations in many shopping experiences, both in online and brick-and-mortar retail, where customers may buy a basket of products instead of a single item. In this paper we incorporate customer's multi-item purchase behavior into the assortment optimization problem. We consider both the uncapacitated and capacitated assortment problems under the so-called Multivariate MNL (MVMNL) model, which is one of the most popular multivariate choice models used in the marketing and empirical literature.

Methodology/results: We first show that the traditional revenue-ordered assortment may not be optimal. Nonetheless, we show that under some mild conditions, a certain variant of this property holds (in the uncapacitated assortment problem) under the MVMNL model---that is, the optimal assortment consists of revenue-ordered local assortments in each product category. Finding the optimal assortment even when there is no interaction among product categories is still computationally expensive as the revenue thresholds for different categories cannot be computed separately. To tackle the computational complexity, we develop FPTAS for several variants of (capacitated and uncapacitated) assortment problems under MVMNL.

Managerial implications: Our analysis reveals that disregarding customer's multi-item purchase behavior in assortment decision can indeed have a significant negative impact on profitability, demonstrating its practical importance in retail. We numerically show that our proposed algorithm can improve a retailer's expected total revenues (compared to a benchmark policy that does not properly take into account the impact of customer's multi-item choice behavior in assortment decision) by up to 14%.

Keywords: Assortment optimization, multi-item choice model, dynamic programming, approximation algorithms