Unified Recommendation and Search in E-Commerce

Abstract

In an e-commerce website, a recommender system and a search engine are usually developed and used separately. In fact they serve a similar goal: helping potential consumers find products to purchase. Recommender systems use a user’s prior purchase history to learn the user’s preferences and recommend products that the user might like. Search engines use a user’s query information, which tells much about a user’s purchase intention in the current search session, to find matching products. This paper explores how to integrate the complementary information to build a unified recommendation and search system. We propose three approaches. The first one is using a multinomial logistic regression model to integrate a rich set of search features and recommendation features. The second one is using a gradient boosted tree based ranking model. The third one is a new model that explicitly models the user’s categorical choice, purchase state (repurchase, variety seeking or new purchase) in addition to the final product choice. Experiments on a data from an e-commerce web site (shop.com) show that unified models work better than the basic search or recommendation systems on average, particularly for the repeated purchase situations. The new model predicts a user’s categorical choice and purchase state reasonably well. The insight and predicted purchase state may be useful for implementing the user-state specific marketing and advertising strategies.