TSL.zip_tsl_概率分布失配_欠适配_欠适配问题_负迁移负适配

function P = TSL_LRSR(Xs,Xt,Xs_label,alpha,beta,lambda) Y = Construct_Y(Xs_label,length(Xs_label)); B = Construct_B(Y); Class = length(unique(Xs_label)); Max_iter = 100; [m,n1] = size(Xs); n2 = size(Xt,2); max_mu = 10^6; mu = 0.1; rho = 1.01; convergence = 10^-6; options = []; options.ReducedDim = Class; [P1,~] = PCA1(Xs', options); % ---------------------------------------------- % Initialization % ---------------------------------------------- M = ones(Class,n1); P = zeros(m,Class); E = zeros(Class,n2); Z = zeros(n1,n2); Z1 = zeros(n1,n2); Z2 = zeros(n1,n2); Y1 = zeros(Class,n2); Y2 = zeros(n1,n2); Y3 = zeros(n1,n2); % ------------------------------------------------ % Main Loop % ------------------------------------------------ for iter = 1:Max_iter % updating P V0 = Y+B.*M; V1 = Xt-Xs*Z; V2 = E-Y1/mu; if (iter == 1) P = P1; else P = (2*Xs*Xs'+mu*V1*V1'+2*lambda*eye(m))\(2*Xs*V0'+mu*V1*V2'); end % updating M R = P'*Xs-Y; gp = B.*R; [numm1,numm2] = size(gp); for jk1 = 1:numm1 for jk2 = 1:numm2 M(jk1,jk2) = max(gp(jk1,jk2),0); end end % updating E the2 = beta/mu; temp_E = P'*Xt-P'*Xs*Z+Y1/mu; E = max(0,temp_E-the2)+min(0,temp_E+the2); % updating Z V3 = Z1-Y2/mu; V4 = Z2-Y3/mu; V5 = P'*Xt-E+Y1/mu; Z = (2*eye(n1)+Xs'*P*P'*Xs)\(V3+V4+Xs'*P*V5); % updating Z1 ta = 1/mu; temp_Z1 = Z+Y2/mu; [U01,S01,V01] = svd(temp_Z1,'econ'); S01 = diag(S01); svp = length(find(S01>ta)); if svp >= 1 S01 = S01(1:svp)-ta; else svp = 1; S01 = 0; end Z1 = U01(:,1:svp)*diag(S01)*V01(:,1:svp)'; % updating Z2 taa = alpha/mu; temp_Z2 = Z+Y3/mu; Z2 = max(0,temp_Z2-taa)+min(0,temp_Z2+taa); % updating Y1, Y2, Y3 Y1 = Y1+mu*(P'*Xt-P'*Xs*Z-E); Y2 = Y2+mu*(Z-Z1); Y3 = Y3+mu*(Z-Z2); % updating mu mu = min(rho*mu,max_mu); % checking convergence leq1 = norm(P'*Xt-P'*Xs*Z-E,Inf); leq2 = norm(Z-Z1,Inf); leq3 = norm(Z-Z2,Inf); if iter > 2 if leq1<convergence && leq2<convergence && leq3<convergence break end end end end function B = Construct_B(Y) %% B = Y; B(Y==0) = -1; end function Y = Construct_Y(gnd,num_l) %% % gnd:标签向量； % num_l:表示有标签样本的数目； % Y:生成的标签矩阵； nClass = length(unique(gnd)); Y = zeros(nClass,length(gnd)); for i = 1:num_l for j = 1:nClass if j == gnd(i) Y(j,i) = 1; end end end end