In E-commerce recommendation system, "Information overload" on Internet has brought a tough problem, which is how to precisely position users' interest and provide users with accurate brand recommendation. To solve this problem, in this paper, many features which could describe the purchasing behavior of users are extracted by deeply mining large-scale of user behavior logs. A brand preference model was constructed by applying these features into Gradient Boosting Regression Tree algorithm, to improve accuracy of the recommendation algorithm. Experiment results show that, in condition of sparse data, algorithm in this paper can still fit brand preference of users very well, and has significantly improvement in accuracy compared with traditional recommendation and classification algorithm.