Retargeted Multicategory Classification Algorithm Based On Random Weighted Neural Network

Classification is one of the most important research fields of Machine Learning,and multi-category classification is a hot issue in the study of classification.Recently,to deal with multi-category classification problems,a discriminative least squares regression(DLSR)algorithm is proposed,in which to rebuild the zero-one target matrix,a technique called?-dragging is integrated into the least squares loss function.This technique forces the regression targets of different classes moving along opposite directions such that the distances between classes can be enlarged.Motivated by rebuilding target matrix,retargeted least squares regression(ReLSR)algorithm is proposed subsequently.Compared with DLSR models,ReLSR learns the regression targets from input data directly,which is different from ?-dragging and can be more efficient in reducing the classification error.However,DLSR and ReLSR adopt linear functions as the classifiers.And the relationship among practical data is much more complicated than linear model.In this brief,we choose nonlinear model,random weighted neural network(RWNN),as classifier and propose a retargeted multiclass algorithm based on random weighted neural network(ReRWNN).Compared with ReLSR,ReRWNN reach more accurate in classification and better ability in processing complex data.