Predicting online user purchase behavior based on browsing history

Recently, people tend to purchase through websites. This change allows e-commerce sites to collect user behavior data from web logs. E-commerce marketing forces usually make use of such data to come up with subsequent promotional campaign to drive more traffic, and converting into paying customers. In this paper we consider a special kind of e-commerce companies which sell products with similar property and usually at a high price. Therefore, the recommendation becomes less important than prediction of items(if any) bought. We want to discover potential buyers and deliver ads or even coupons to them, expecting them to be real buyers. In this paper, we model the buying behaviors from clicking records with patterns extracted using feature engineering approach. Our solution was to model two kinds of browsing behaviors, namely hesitant and impulsive respectively. In the model, we define some interaction features from click-streams which uncover users' purchase intention with the product pages, how long the user stays on that page, and then build a model which can predict users' preference. Experimental results on a real dataset from an e-commerce company demonstrate the effectiveness of the proposed method. The approaches in our work can be used to model user purchasing intent and applied to e-commerce sites which sell high-end products.