Optimization of Recommender Systems Based on Inventory

Abstract

We consider subscription‐based rental organizations, such as Netflix, where the satisfaction of customers depends on the availability of requested products. Recommender systems, in a DVD‐rental context, are typically used to help customers in finding the right movies for them. Accordingly, the focus in recommender system research is generally on making better predictions of users' ratings. In contrast, we focus on better utilizing these rating estimates in the operations of DVD‐rental firms. We show that a more explicit consideration of inventory level and future demand can help the firms better manage demand, and can increase the number of satisfied customers substantially. However, if the uncertainty regarding inventory levels is high, the performance of one of our proposed approaches may be worse than the prevalent industry practice under certain conditions. We discuss these conditions and propose a quick recipe for dealing with high levels of variation in inventory estimation. We show that when it is not possible to estimate inventory levels reliably, it is better to underestimate rather than overestimate. Other findings include the trade‐off between the short‐term profitability of the firms and long‐term customer trust; and the effect of variation in rating estimates on the quality of our solution approaches.