PINSVR

A Matlab code for parametric-insensitive nonparallel support vector machine for regression. [Code]

Reference

Zhi-Min Yang, Xiang-Yu Hua, Yuan-Hai Shao*, Ya-Fen Ye. A novel parametric-insensitive nonparallel support vector machine for regression, Submitted,2014.

Main Function

function [PredictY,PredictYu,PredictYd]=PINSVR(TestX,X,Y,c1,c2,c3,c4,kersign,p) %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % PINSVR: A novel parametric-insensitive nonparallel support vector machine for regression % [PredictY,PredictYu,PredictYd]=PINSVR(TestX,X,Y,c1,c2,c3,c4,kersign,p) % % Input: % TestX - Test sample matrix. Each row vector of fea is a test sample. % X - Training sample matrix. Each row vector of fea is a training sample. % Y - Training sample vector. Each element is a value of a training sample. % c1 - [0,inf] Paramter to tune the weight. % c2 - [0,inf] Paramter to tune the weight. % c3 - [0,inf] Paramter to tune the weight. % c4 - [0,inf] Paramter to tune the weight. % kersign - kernel type. % p - kernel parameter. % % Output: % Predict_Y - Predict value of the TestX. % PredictYu - Predict up-bound value of the TestX. % PredictYd - Predict down-bound value of the TestX. % % Example: % load ExamIA.mat X Y TestX TestY % c1=2^(-1); % c2=2^(-1); % c3=0.01; % c4=0.01; % kersign='rbf'; % p=3; % [PredictY,PredictYu,PredictYd]=PINSVR(TestX,X,Y,c1,c2,c3,c4,kersign,p); % % Reference: Zhi-Min Yang, Xiang-Yu Hua, Yuan-Hai Shao, Ya-Fen Ye, % A novel parametric-insensitive nonparallel support vector machine for regression, % Submitted, 2014. % % Version 1.0 - Sep/2014 % % Written by Xiang-Yu Hua (hxy9069@163.com) %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Initailization %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %tic; kerfPara.type=kersign; kerfPara.pars=p; epsilon=1e-10; [m,n]=size(X); e1=ones(m,1); %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Compute Kernel %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% if strcmp(kerfPara.type,'lin') G=[X e1]; I1=eye(n+1); I2=eye(m); else G=[kernelfun(X,kerfPara,X) e1]; I1=eye(m+1); I2=eye(m); end %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Train Samples using QP solver %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% U1=(G'*G+c3*I1)\G'; U2=(G'*G+c4*I1)\G'; QPH01=[1/c3*(G*G') 1/c3*(G*G');1/c3*(G*G') (G*U1+1/c3*(G*G'))]; QPH02=[1/c4*(G*G') 1/c4*(G*G');1/c4*(G*G') (G*U2+1/c4*(G*G'))]; QPH01=QPH01+epsilon*eye(size(QPH01)); QPH02=QPH02+epsilon*eye(size(QPH02)); QPH1=(QPH01+QPH01')/2; QPH2=(QPH02+QPH02')/2; clear QPH01 QPH02 QPf1=[zeros(m,1);-(Y'*(G*U1-I2))']; QPf2=[zeros(m,1);(Y'*(G*U2-I2))']; QPA1=[];QPb1=[];QPB1=[];QPd1=[]; QPvlb1=zeros(2*m,1);QPvub1=[inf*ones(m,1);c1*ones(m,1)]; QPA2=[];QPb2=[];QPB2=[];QPd2=[]; QPvlb2=zeros(2*m,1);QPvub2=[inf*ones(m,1);c2*ones(m,1)]; QPx1=quadprog(QPH1,QPf1,QPA1,QPb1,QPB1,QPd1,QPvlb1,QPvub1); QPx2=quadprog(QPH2,QPf2,QPA2,QPb2,QPB2,QPd2,QPvlb2,QPvub2); %x=[alpha;beta] alpha1=QPx1(1:m);beta1=QPx1((m+1):2*m); alpha2=QPx2(1:m);beta2=QPx2((m+1):2*m); u1=U1*(Y-beta1); u3=1/c3*G'*(alpha1+beta1); u2=U2*(Y+beta2); u4=1/c4*G'*(alpha2+beta2); clear G %u=[w;b] k=size(u1,1); w1=u1(1:(k-1));b1=u1(k); w2=u2(1:(k-1));b2=u2(k); w3=u3(1:(k-1));b3=u3(k); w4=u4(1:(k-1));b4=u4(k); %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % Predict and output %%%%%%%%%%%%%%%%%%%%%%%%%%%%%% e2=ones(size(TestX,1),1); if strcmp(kerfPara.type,'lin') f1=TestX*w1+b1*e2; f2=TestX*w2+b2*e2; g1=TestX*w3+b3*e2; g2=TestX*w4+b4*e2; else H=kernelfun(TestX,kerfPara,X); f1=H*w1+b1*e2; f2=H*w2+b2*e2; g1=H*w3+b3*e2; g2=H*w4+b4*e2; end PredictY=0.5*(f1+f2); PredictYu=f2+g2; PredictYd=f1-g1; end

PINSVR

Reference

Main Function

Contacts