Multiple Regression Method Based on Unexpandable and Irreducible Convex Combinations

A new multiple regression method based on optimal convex combinations of simple univariate regressions is discussed, where the simple regressions are searched with an ordinary least squares technique. Convex combination is considered optimal if it correlates with the response variable in the best way. It is shown that the developed approach is equivalent to a least squares technique variant regularized by constraints on signs of regression parameters. A method of optimal convex combination search is discussed that is based on unexpandable and irreducible ensembles. The developed method is compared with elastic networks at simulated data.