A Multiarmed Bandit Approach for House Ads Recommendations.

A multiarmed bandit approach that uses a deep neural network to decide personalized recommendations for display in the homepage of an online retailer. Nowadays, websites use a variety of recommendation systems to decide the content to display to their visitors. In this work, we use a multiarmed bandit approach to dynamically select the combination of house ads to exhibit to a heterogeneous set of customers visiting the website of a large retailer. House ads correspond to promotional information displayed on the website to highlight some specific products and are an important marketing tool for online retailers. As the number of clicks they receive not only depends on their own attractiveness but also on how attractive are other products displayed around them, we decide about complete collections of ads that capture those interactions. Moreover, as ads can wear out, in our recommendations we allow for nonstationary rewards. Furthermore, considering the sparsity of customer-level information, we embed a deep neural network to provide personalized recommendations within a bandit scheme. We tested our methods in controlled experiments where we compared them against decisions made by an experienced team of managers and the recommendations of a variety of other bandit policies. Our results show a more active exploration of the decision space and a significant increment in click-through and add-to-cart rates. History: Olivier Toubia served as the senior editor and Hema Yoganarasimhan served as associate editor for this article. Funding: This work was supported by the Institute for Research in Market Imperfections and Public Policy [Grant ICM IS130002]. M. Schiappacasse gratefully acknowledges partial funding by the Institute for Research in Market Imperfections and Public Policy (IS130002 ANID). Supplemental Material: The data files and online appendix are available at https://doi.org/10.1287/mksc.2022.1378. [ABSTRACT FROM AUTHOR]