E-Commerce Product Recommendation Using Historical Purchases and Clickstream Data

Abstract

In E-commerce, user-item rating matrices for collaborative filtering recommendation systems are usually binary and sparse, showing only whether or not a user has purchased an item previously. Clickstream data containing more customer behavior have been used to improve recommendations by some existing systems referred in this paper as Kim05Rec, Kim11Rec, and Chen13Rec, using decision tree, association rule mining and category-based interest measurements respectively. However, they do not integrate valuable information from historical purchases and the consequential bond information between session-based clicks and purchases. This paper proposes Historical Purchase with Clickstream recommendation system (HPCRec), which normalizes the historical purchase frequency matrix to improve rating quality, and mines the session-based consequential bond between clicks and purchases to generate potential ratings to improve the rating quantity. Experimental results show HPCRec outperforms these existing methods, and is also capable of handling infrequent user cases, whereas other methods can not.