Tensor Time Series Analysis for LTV Prediction

Abstract

Lifetime Value (LTV) prediction is a crucial problem for customer evaluation, where an accurate estimate of future value allows retailers to realize any successful customer relationship management strategy. Given a large purchase history, which consists of multiple attributes such as timestamp, product category, and user ID, how can we find underlying patterns and trends? How accurately can we predict user activities and their LTVs? In this paper, we propose a novel way to predict LTV, which performs multi-way mining to discover hidden topics, groups of products, and groups of users, simultaneously. The comprehensive summarization makes it possible to accomplish LTV prediction with user characteristics learned from data. Experiments on real datasets demonstrate the benefits of the proposed model, in that the model can capture interpretable topics across all aspects of the purchase history and outperforms its baseline methods.