Some Learning Procedures for Binary and Multiple Classification Problems

Abstract

The projection uLSIF (PuLSIF) and projection uLSIF rotation data(PuLSIF_RD) are the improvements of unconstrained least-square importance fitting (uLSIF) We apply these ideas which try to the projection and rotation subspace to improve the least-square probabilistic classification (LSPC) However the improvement is not significant and time consuming is much longer Moreover we are informd that the naive linear regression method can be linked the Fisher’s discriminant analysis when binary classes We imagine that samples are projected along the line constructed by the linear regression method and allocate the samples If such cases as the training sample are expensive or the sample size are extremely large are encountered And under this active learning scenario the sequential method is a nature technique In short we use the iteration to reduce the training sample size Thus in this paper there will be presented some criteria with the regression discriminant classifier and these criteria can economize on training samples Besides we do not focus on the modeling part but the predition part As long as two performances of small training sample and all training sample size are close then solve the problem about training cost Indeed its result can be involved in big data issue Numerical result can conclude that these strategies are comparable to the all sample size into regression discriminant classifier in accuracy even AUC part

Date of Award	2015 Aug 12
Original language	English
Supervisor	Ray-Bing Chen (Supervisor)