CMStatistics 2016: Start Registration
View Submission - CMStatistics
B1728
Title: Primal path algorithm for compositional data analysis Authors:  Jong-june Jeon - University of Seoul (Korea, South) [presenting]
Abstract: The compositional data has two distinguished characteristics compared to a general multivariate data: the values in the observations are nonnegative; the summand of the values is exactly 1. We propose an efficient path algorithm of lasso for the analysis of compositional data. The proposed algorithm has three advantages over the previously developed algorithms. First the algorithm gives the exact solution path based on checking the Karush-Kuhn-Tucker conditions of convex function with linear constraints. Second, the algorithm is easy to extend to the regularized regression problem with general loss functions such as Huberized loss. Third, the algorithm gives an exact ordinary least square estimator in the end of optimizations. We also develop the classification model for the compositional data with the proposed algorithm.