A Large-Scale Marketing Model using Variational Bayes Inference for Sparse Transaction Data

Large-scale databases in marketing track multiple consumers across multiple product categories. A challenge in modeling these data is the resulting size of the data matrix, which often has thousands of consumers and thousands of choice alternatives with prices and merchandising variables changing over time. We develop a heterogeneous topic model for these data, and employ variational Bayes techniques for estimation that are shown to be accurate in a Monte Carlo simulation study. We find the model to be highly scalable and useful for identifying effective marketing variables for different consumers, and for predicting the choices of infrequent purchasers.