When managing multiple stores in the same marketplace, retailers need to select store locations and localize product assortments to reflect the heterogeneous demand preferences across communities. This paper develops a dual Poisson Dynamic System with Multilayer Factorization (dPDS-MF) for panel data on product assortments and individual consumers’ purchases across store/vending locations. The dPDS-MF can help retailers automatically profile different consumer segments driven by store visiting preferences, measure the relationships across store locations, and estimate the product preferences for each consumer segment simultaneously. The dPDS-MF relies on a Bayesian nonparametric prior and can be efficiently trained for large-scale transactional data across hundreds of stores and SKUs, using our proposed MCMC inference algorithm. We apply the dPDS-MF in the retail vending market in major train stations in Japan. We demonstrate the face validity of the direct outputs from the dPDS-MF for improving vending location decisions as well as location-specific assortments. More importantly, we showcase how the dPDS-MF can be combined with a choice model to solve the optimal localized assortments efficiently and effectively. We show that compared with several benchmark strategies, including the nested-logit choice model, our proposed assortment strategy not only improves the expected revenue up-to 30% but also gives more meaningful localized assortment decisions.
Keywords: product assortments, consumer segmentation, topic modeling, choice modeling, optimization, big data, decision analytics