中国科学院自动化研究所机构知识库

In this paper, we report the solution of our team ”Farewell” on the IJCAI 15 competition. This com- petition provides a set of merchants and their cor- responding new buyers acquired during the promo- tion on the ”Double 11” Day and the task is to predict which users will become loyal customers for given merchants in the future. To solve this problem, we firstly do a lot of work in feature engineering. Features are extracted from various aspects and divided into several categories. Be- sides, we adopt some innovative methods to gen- erate latent factor features and user-merchant co- cluster features. In the first stage of the compe- tition, we employ two individual models, Factor- ization Machines(FM) and Gradient Boosting De- cision Tree(GBDT) and we adopt Logistic Regres- sion(LR) model in the second stage. Afterwards, we propose a weighted linear ensemble model to combine results of individual models. Our final re- sult is the blending of FM model and GBDT model in stage 1 and result of LR model in stage 2.