Purchase Prediction Using Tmall-specific Features

Yang Zhao,Liang Yao,Yin Zhang

DOI: https://doi.org/10.1002/cpe.3720

2016-01-01

Concurrency and Computation Practice and Experience

Abstract:SummaryHistorical user activity, such as online shopping recommendations, content personalization, and advertising clicking rates, is a critical component of building user profiles to predict purchases and preferences. Alongside the rapid development of e‐commerce, purchase prediction has become an increasingly important consideration for a wide variety of retail platforms. This paper proposes a framework which combines machine learning methods with a threshold‐moving approach to predict sets of pairs (user id and brand id,) in terms of whether a certain brand is purchased by a specified user according to his or her historical activity records. Three specific feature groups are extracted: click features, purchase features, and collect‐and‐cart features using a dataset from Tmall, a Chinese business‐to‐consumer online retail platform. Next, seven user purchase prediction experiments with different combinations of the three feature groups are conducted, and the purchase prediction performance is observed. Results showed that a combination of all three feature groups, with 27 features in total, provided valuable purchase prediction contributions and performed favorably. Also, three feature groups are proved to be relative independent, and our prediction model identified the most important feature, from the collect‐and‐cart feature group. Detailed result analysis validated the effectiveness of both the extracted features and proposed machine learning methods. Copyright © 2016 John Wiley & Sons, Ltd.