Marketing Analytics for High-Dimensional Assortments Defended on Friday, 22 December 2017
Over the past two decades online retailing has become ubiquitous and today’s large online retailers enable customers to purchase virtually any product. As a consequence product assortments at such retailers are of a different order of magnitude compared to the traditional brick-and-mortar stores. In this dissertation model-based methods are presented that can be used to model purchase decisions in such high-dimensional product assortments. These methods are able to accurately predict at the individual customer level which product will be purchased next out of the large assortment. In addition, the methods provide substantive insights in the patterns that underlie the observed purchase behavior. The applicability of such methods in practice hinges on their scalability and this holds especially true for online retailers. Model results should be rapidly obtained and the estimation time should not significantly increase in case the customer base or product assortment expands. Scalability is therefore a focal point in this dissertation. The methods introduced are adaptations and extensions of fast scalable methods from the machine learning literature that make these methods also suitable for the online retailing context. This ensures that estimation times remain feasible even if the size of the retailer increases and opens the way for advanced model-based marketing analytics in high-dimensional assortments.
Marketing, high-dimensional assortments, model-based predictions, large scale purchase prediction, marketing analytics, scalability, purchase history data, latent Dirichlet allocation, variational inference,