forked from huangsuj/ERL-MVSC
-
Notifications
You must be signed in to change notification settings - Fork 0
/
MVAR.m
156 lines (133 loc) · 4.54 KB
/
MVAR.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
function [W,b,F,ypre] = MVAR(Xl,Xu,Yl,lambda, s, r,maxIter)
%% code of "Scalable Multi-View Semi-Supervised Classification via Adaptive Regression"
%% Tao et al. IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 26, NO. 9, SEPTEMBER 2017
%% input:
%%%% Xl & Xu: cell, each cell element is a (labeled or unlabeled) data
%%%% matrix from one view, each row is a data point.
%%%% the data matrix may be pre-processed by data normalization
%%%% or centerization when necessary
%%%% Yl: label matirx of the labeled part, 1-of-c coding, i.e., Y(i,j) = 1, if i belongs to
%%%% the j-th class, otherwise Y(i,j) = 0.
%%%% s: array, the weights manually assigned for each points,usually the weights
%%%% for labeled points is larger than unlabeled points.
%%%% lambda: array, the trade-off parameters for each view
%%%% r>1: the parameter for redistribute weights over views
if ~exist('maxIter','var')
maxIter = 30;
end
s = s(:);
viewNum = length(Xl);
[nlSmp,nClass] = size(Yl);
nuSmp = size(Xu{1},1);
nSmp = nlSmp + nuSmp;
viewDim = zeros(viewNum,1);
%% initialization
alpha = ones(viewNum,1)/viewNum; %% initialize the view weight
W = cell(viewNum,1); b = W;
for v = 1:viewNum
viewDim(v) = size(Xl{v},2);
[W{v}, b{v}] = least_squares_regression(Xl{v}, Yl, lambda(v));
end
%% parameters for stopping the loop
h = 6;
do_loop = 1;
obj = [];
stop_crit = 1e-6;
%% b is absorbed into W
X = cell(viewNum,1);
XX = cell(viewNum,1); WW = XX;
en = ones(nSmp,1);
B= cell(viewNum,1); bb = B;
for v = 1:viewNum
X{v} = [Xl{v}; Xu{v}];
XX{v} = [X{v} en];
WW{v} = [W{v}; b{v}'];
B{v} = diag(ones(nSmp,1));
end
%% begin the loop
iter = 1;
F = zeros(nSmp,nClass);
F(1:nlSmp,:) = Yl;
while iter < maxIter && do_loop
%%%% update F
Fu = zeros(nuSmp,nClass);
for v = 1:viewNum
temp = Xu{v}*W{v} +ones(nuSmp,1)*b{v}';
Fu = Fu + alpha(v)^r*(B{v}(nlSmp+1:nSmp,nlSmp+1:nSmp)*temp);
end
[~,ypre] = max(Fu,[],2);
Ypre = zeros(nuSmp,nClass);
for i = 1:nClass
Ypre(ypre == i,i) = 1;
end
F(nlSmp+1:nSmp,:) = Ypre;
%%%% update matrix B
for v = 1:viewNum
Ev = XX{v}*WW{v} - F;
bb{v} = 0.5*s./sqrt(sum(Ev.*Ev,2) + eps); %%%% eps is added to avoid being divided by 0.
B{v} = diag(bb{v});
end
%%%% update WW
%%%% when the matrix G is close to singularity, the results may be not
%%%% precise, this will also affect the convergence of objective functioin
for v = 1:viewNum
if viewDim(v) < nSmp
G = XX{v}'*B{v}*XX{v} + lambda(v)*eye(viewDim(v)+1);
WW{v} = G\(XX{v}'*B{v}*F);
else
G = XX{v}*XX{v}' + lambda(v)*diag(1./bb{v});
WW{v} = XX{v}'*(G\F);
end
W{v} = WW{v}(1:end-1,:);
b{v} = WW{v}(end,:)';
end
%%%% update alpha and calculate objective value
resErr = zeros(viewNum,1);
for v = 1:viewNum
ErrMat = XX{v}*WW{v} - F;
resErr(v) = sum(s'*sqrt(sum(ErrMat.*ErrMat,2))) + lambda(v)*sum(sum(WW{v}.*WW{v}));
end
if r > 1
alpha = resErr.^(1/(1-r));
alpha = alpha/sum(alpha);
else
error('r must larger than 1.\n');
end
if iter > 1
obj(iter-1) = sum((alpha.^r).*resErr);
end
if iter > h
temp = obj(end-h+1:end);
objdiff(iter-h+1) = (max(temp) - min(temp))/max(temp);
end
if exist('res','var') && objdiff(end) < stop_crit
do_loop = 0;
end
iter = iter + 1;
end
end
function [W, b] = least_squares_regression(X, Y, gamma)
% X: each row is a data point
% Y: each row is an target data point: such as [0, 1, 0, ..., 0]'
% gamma: a positive scalar
[N, dim] = size (X);
[~, dim_reduced] = size(Y);
% first step, remove the mean!
XMean = mean(X);
XX = X - repmat(XMean, N, 1);
W = [];
b = [];
if dim < N
% W = pinv( XX * XX' + gamma * eye(dim)) * (XX * Y');
% Note that the above sentence can be repalced by the following sentences. So, it is more fast.
t0 = XX' * XX + gamma * eye(dim);
W = t0 \ (XX' * Y);
b = Y - X*W;
b = mean(b)';
else
t0 = XX * XX' + gamma * eye(N);
W = XX' * (t0 \ Y);
b = Y - X*W;
b = mean(b)';
end
end