Model-based Purchase Predictions for Large Assortments

Being able to accurately predict what a customer will purchase next is of paramount importance to successful online retailing. Given the large assortments maintained by online retailers, scalability of the prediction method is just as important as its accuracy. In this paper we study scalable model-based purchase predictions and contrast their performance with that of collaborative filtering algorithms. Such algorithms are commonly used in practice. We consider two classes of models: mixtures of Dirichlet-Multinomials (MDM), which have a rich history in brand choice modeling, and latent motivation models (LMM(-X)), which extend latent Dirichlet allocation (LDA). We illustrate a key benefit of model-based approaches, which is their ability to accommodate observed customer heterogeneity through the inclusion of predictor variables. The prediction methods are compared based on their heterogeneity assumptions and scalability in terms of estimation com- plexity and in terms of memory requirements for online predictions. We apply the methods to purchase data from an online retailer and show that the model- based approaches consistently outperform the collaborative filtering algorithms. Moreover, LMM(-X) attains performance similar to that of MDM, while being more scalable in terms of assortment size, rendering the latent motivation models a promising model-based approach to purchase prediction in large assortments.