Forecast customer flow using long short-term memory networks

Abstract

Customer flow forecast is of practical importance in business intelligence domain. This paper particularly investigates an interesting issue, i.e., how to forecast off-line customer flow for over two thousand shops by considering both online customer behaviors and off-line periodic customer behaviors. Apparently, it is difficult to directly model these underlying dependent variables via traditional regression models. To this end, the proposed approach first introduces various extra information to incorporate more underlying factors. Then, the hierarchical linear model is performed to screen out insignificant factors. On the basis of this reduced feature space, the second-order flow factor is incorporated to model the variance term. The combined new feature set is then used for the learning of a number of Long Short Term Memory (LSTM) models. The rigorous experiments have been performed and the promising results demonstrate the superiority of the proposed approach which indicates the wide applicability of the proposed forecast model.