Predicting Purchase Behavior of E-commerce Customer, One-stage or Two-stage?

Abstract

The development of Internet has led to huge economic benefits and challenges for e-commerce companies, which develop online shopping stores to serve online customers, as well as collect valuable data of customers' behavior. Purchase behavior prediction is one of the most important issues to promote both companies' sales and consumers' satisfaction. This paper studies two modeling strategies of purchase behavior prediction in the engineering practice: one-stage and two-stage. The appropriate modeling strategy is essential to implement an effective and efficient predictive model. The pros and cons of one-stage and two-stage modeling is analyzed and discussed. We also conduct experiments on real-world and large-scale dataset and find that the performance of two-stage modeling is usually better than one-stage. The reason is that two-stage modeling can make better use of the feature engineering for each stage and reduce the requirement for machine learning algorithms. The purchase behavior can be better predicted by the machine learning algorithm which is ensemble and able to use high-order interactions among features.