A Bayesian framework for large-scale geo-demand estimation in on-line retailing

Time-specific geo-demand distribution estimation of the products in the catalog is the fundamental guiding analytics for inventory allocation in any major online retailer’s supply chain operations. Although geography-specific historical sales data is available for learning the geo-demand distributions, it is extremely sparse from a view of a product × demand zone × time data cube (tensor). As a result, we have to estimate the entries in a large-scale tensor with limited amount of training data. The sheer scale of the problem makes the task challenging to solve within a limited time frame. We formulate this problem in the spirit of text theme classification and view the geo-demand distributions as underlying probability distributions that govern the historical sales observations. We develop a Bayesian framework based on mixture of Multinomials for estimating the time-dependent geo-demand distributions in a collaborative manner. As a by-product, the solution provides guidance on grouping the products by their geo-demand patterns. We also provide practical solutions to counter various scalability issues. Benchmark results are provided in comparison to basic same-class methods and a state-of-the-art R package.