Tensor train neural networks in retail operations

Abstract

The neural network generalization of Tensor Train decomposition for multidimensional datasets of censored Poisson counts is presented. The model is successfully applied to two important classes of retail operations: sales process under the controlled stock distribution over the retail network, and the optimization of active retailer decisions, such as pricing policy, marketing actions, and discounts. The advantage of proposed Tensor Train Neural Network model is in its ability to capture non-linear relations between similar retail stores and similar consumer goods, as well as jointly estimate sales potential of commodities with wide dynamic range of popularity.