Abstract: A new regression method based on convex correcting procedures over sets of predictors is developed. In contrast to previously developed approach based on minimization of generalized error, the proposed one utilies correcting procedures of maximal correlation with the target value. In the proposed approach a concept of a set of predictors irreducible against target functional is used where irreducibility is understood as lack of combinations of at least the same value of the functional after removing any of its predictors. Sets of combinations simultaniously irreducilbe and unexpandable are used during the construction of a prognostic rule. Results of some computational experiments described in the present article show an efficiency comparison between the two approaches.

Keywords: forecasting, bias-variance decomposition, convex combinations, variables selection.

ACM Classification Keywords: G.3 Probability and Statistics - Correlation and regression analysis, Statistical computing.

Link:

CORRELATION MAXIMIZATION IN REGRESSION MODELS BASED ON CONVEX COMBINATIONS

Oleg Senko, Alexander Dokukin

http://www.foibg.com/ijita/vol18/ijita18-3-p03.pdf