Yun-Rui Li¹, Ting-Kai Hwang² and Shi-Chung Chang¹, ¹National Taiwan University, Taiwan and ²Ming Chuan University, Taiwan

With more and more people shopping online, companies deal with customer data input not only in high volume but also dynamic. In order to attract target customers more effectively and to provide customers with more personalized services, how to automatically extract personal preference from the real-time data and make real-time recommendation has been growing in importance for businesses in the competitive modern society. Current data analysis methods for online shopping recommendation largely rely on historical transaction record. Analyses have indicating that next items a customer would like to buy not only depend on one’s past historical records but on the item currently being put into the shopping cart. This paper designs an engine to combine each customer’s past transaction and current shopping cart data to dynamically infer one’s preference for the next items. The design, Transaction-Data Based Real-time Preference Inference Engine (TRPIE), consists of two innovative ideas. The first exploits the purchasing sequence information and turns one’s purchase history into a temporal series of data, where a customer’s dynamic purchasing behaviour information lies. The second is a design of a two-layer Recurrent Neural Network (RNN) for extracting personal purchasing preference pattern from the temporal series of data to infer preference of next items. A reference implementation of TRPIE design integrates existing tools such as Keras, tensorflowTM, sklearnTM, and MlxtendTM. Test results over real data from 1,374 people show that prediction accuracy has doubled that obtained by a basket analysis method, which ignores sequentiality of purchasing items.

Online shopping, transaction data,purchasing sequence, dynamic preference inference, re-current neural network.