Research and Implementation of Multiple Behavior Based Recommender System in E-Commerce

Recommender systems provide users with personalized items. These systems often rely on collaborative filtering. Collaborative filtering does recommendation by finding out similar users or items. Collaborative filtering has a fatal weakness that the reliability of the similarity of users or items depends on users' preference list. If most users' preference list is short, namely most users show interest in only a few items, collaborative filtering loses its accuracy in finding similar users or items and performs poorly in recommendation. This defect is particularly obvious in EC system. In this paper, we introduce two regression based methods, logistic regression and gradient boosting regression tree to build a recommender system. Unlike traditional collaborative filtering, regression based methods do not focus on each user or item, but use all users' historical behavior data to build a single model. The single model then produces the probability of items a user might buy according to users' new behavior data. Items with high probability will be recommended to users. Our experiment shows that regression based methods performs much better than collaborative filtering.