CMStatistics 2018: Start Registration
View Submission - CMStatistics
B1760
Title: Multi-component ridge regression for heterogeneous correlation structure of covariates Authors:  Junghwan Kim - Institute of Water Resources System, Inha University (Korea, South) [presenting]
Woojoo Lee - Inha University (Korea, South)
Abstract: Ridge regression is a classical regression method applicable to the data showing high multicollinearity. In practice, one-dimensional tuning parameter is often used for ridge regression. However, in high-dimensional data, blockwise heterogeneous correlation structures among predictors are often observed so that one tuning parameter may not be sufficient to control different degree of multicollinearity simultaneously. We propose a multi-component ridge regression for doing two folds: (1) it automatically finds the correlation blocks of predictors and (2) it allows multi-dimensional tuning parameters. We compare the predictive performance of the proposed multi-component ridge regression with traditional ridge regression through a numerical study and real data analysis.