A Two-Stage Method for Constructing Linear Regressions Using Optimal Convex Combinations

Abstract

Multilevel learning systems have become more popular in pattern recognition and regression analysis. In this paper, a two-level method for constructing a multidimensional regression model is considered, in which a family of optimal convex combinations of simple one-dimensional least-square regressions is generated at the first level. The second level of the proposed learning system is given by an elastic net. Experimental verification presented demonstrate the efficiency of the proposed regression estimation method as applied to problems with a small amount of data.