eXtreme Gradient Boosting for Identifying Individual Users Across Different Digital Devices

摘要

With the increasing popularity of tablets, smartphones and other mobile electronic devices, it is not uncommon for users to complete online tasks through different electronic devices. Identifying individual users across different digital devices is now becoming a hot research topic. Methods based on name, email and other demographic information have received much attention. However, it is often difficult to obtain a complete set of information. In this paper, we use a probabilistic approach for cross-device identity issue and focus on comparing different algorithms. We conduct an in-depth study and expand the attribute of data through the study of the relationship between attributes. Dummy variables are introduced to improve the efficiency of the models. Experimental results on four datasets (released by ICDM Challenge) show that the eXtreme Gradient Boosting can consistently and significantly outperform other algorithms on both accuracy and F1-score. It also consistently provides a better performance compared to the methods we used in ICDM Challenge (We took part in the ICDM 2015 Challenge, and achieved a moderate score ranking use C4.5 and BP model), and achieves a better comprehensive evaluation ranking.