Statistics Seminar（2015-11）

Topic：Tree-based model for high-dimensional prediction and variable selection

Speaker：Ruqoing Zhu, Yale University

Time：Tuesday,23th, June,14:30-15:30

Location：K01 of Guanghua Hotel

Abstract: Ensemble tree-based methods, such as random forests, are among the state-of-the-art machine learning tools for classication, regression and other statistical modeling problems. However, some intrinsic mechanisms in the tree construction processes have limited the performance of tree-based methods in the "small-n-large-p" paradigm. Reinforcement learning trees (RLT) is introduced to overcome these limitations and to achieve better prediction and variable selection performances. This new method implements an extremely greedy splitting rule to pursue signals. A variable muting procedure is also proposed such that the constructed trees are much sparser than those in traditional tree-based methods under high-dimensional settings. The asymptotic properties of the proposed method are investigated. We discuss the potential applications of this method in cancer genetic studies, personalized medicine, and other related fields.

Your participation is warmly welcomed!