Assortment Optimization under the Multinomial Logit Model with Covering Constraints

TL;DR

This paper develops approximation algorithms for assortment optimization under the multinomial logit model with covering constraints, addressing both deterministic and randomized cases, and validates findings with real-world data.

Contribution

It introduces new approximation algorithms for deterministic assortment optimization with covering constraints and shows polynomial-time solvability for the randomized case.

Findings

Deterministic approximation algorithm with ratio 1/(log K+2)

Polynomial-time solution for the randomized problem

Practical feasibility of covering constraints with minimal revenue loss

Abstract

We consider an assortment optimization problem under the multinomial logit choice model with general covering constraints. In this problem, the seller offers an assortment that should contain a minimum number of products from multiple categories. We refer to these constraints as covering constraints. Such constraints are common in practice due to service level agreements with suppliers or diversity considerations within the assortment. We consider both the deterministic version, where the seller decides on a single assortment, and the randomized version, where they choose a distribution over assortments. In the deterministic case, we provide a $1/(\log K+2)$-approximation algorithm, where $K$ is the number of product categories, matching the problem's hardness up to a constant factor. For the randomized setting, we show that the problem is solvable in polynomial time via an equivalent linear program. We also extend our analysis to multi-segment assortment optimization with covering constraints, where there are $m$ customer segments, and an assortment is offered to each. In the randomized setting, the problem remains polynomially solvable. In the deterministic setting, we design a $(1 - \epsilon) / (\log K + 2)$-approximation algorithm for constant $m$ and a $1 / (m (\log K + 2))$-approximation for general $m$, which matches the hardness up to a logarithmic factor. Finally, we conduct a numerical experiment using real data from an online electronics store, categorizing products by price range and brand. Our findings demonstrate that, in practice, it is feasible to enforce a minimum number of representatives from each category while incurring a relatively small revenue loss. Moreover, we observe that the optimal expected revenue in both deterministic and randomized settings is often comparable, and the optimal solution in the randomized setting typically involves only a few assortments.