Improving Greedy Algorithms for Assortment Optimization Problems with Machine Learning

Assortment optimization is an important problem that arises in many practical applications such as retailing and online advertising.  In this paper, we propose a new data-driven approach to assortment optimization that leverages recent advances in learning machine learning pipelines with combinatorial optimization layers. Under the MNL model, the optimal assortment is revenue ordered. Instead, our approach can be interpreted as a smart greedy algorithm that returns a score ordered assortment, where the scores are learned from data. Since the scores are learned offline, our approach is as fast as a greedy algorithm. Moreover, we conduct extensive numerics and show that it returns near optimal solutions for a variety of choice models and constrained settings.