A Time-aware Multi-task Learning Model for Customer Value Prediction in Civil Aviation

Abstract

The precise prediction of customer value is essential for any successful dynamic customer relationship management (CRM) system. It is also the key for the company to maximizing customer returns. In this research, we concentrate on two main aspects of the work in civil aviation field. Firstly, a reasonable value model is the premise of this prediction issue. Therefore, we propose a parametric customer value model RFUM to estimate customer value in civil aviation. It evaluates customer value from four different attributes and then presents customer value by the weight of the attributes. Secondly, Time-aware Multi-task Value Prediction (TMVP) model is proposed to predict the future value of customer. It employs two supervisory signals of purchase propensity and customer value to better train a specific neural network to automatically learn features. Experiments demonstrate that the RFUM model can more accurately measure the value of customer in civil aviation market and the TMVP model can achieve a more precise regression prediction result. In addition, we also find that increasing the time of a single calculation window can improve the performance markedly.