mddemo.m

偏最小二乘算法在MATLAB中的实现

echo on, clc
%MDDEMO Demonstrates MDPCA for PCA with missing data 

% This script demos the mdpca function.  It takes the pca data 
% from the PLS_Toolbox and leaves out 1% of the data at random.  
% Results can be compared to using pca on the orignal data 
% with nothing left out.

% The pcadata will be loaded and mean centered.
pause

echo off
%  Copyright
%  Barry M. Wise
%  1992
%  Modified May 1994
rand('uniform')
load pcadata
mpart1 = mncn(part1/10);
[m,n] = size(mpart1);
echo on

% Now we can use the pca function to do pca the usual 
% way on the original data.  The scores and loading will be 
% calculated but the plots will be supressed.
pause

[oscores,oloads,ssq,res,q,tsq] = pca(mpart1,0,1:m,4);
pause

% Now 1% of the data will be left out at random 
% and replaced with the flag 9999.

echo off
x = rand(m,n);
missing = [];
for i = 1:m
  for j = 1:n
    if x(i,j) > .99
      mpart1(i,j) = 9999;
      missing = [missing; [i j]];
    end
  end
end
pause
echo on
% We'll now use the mdpca function on the data.  The function 
% will iterate until the estimated values of the missing data 
% converge.  We'll use a tolerance of .001 rather than the 
% default of .00001.
pause

[scores,loads,estdata] = mdpca(mpart1,4,9999,.001);
pause

% We can compare the scores and loadings by plotting them on top
% of each other.
pause
echo off
plot(scores,'+'), hold on, plot(oscores,'-'), hold off
title('Original scores (solid) and estimated scores (+)')
xlabel('Sample Number')
ylabel('Score')
pause
echo on

% Note some scores might be negative of originals.  We can also 
% compare the loadings vectors.
echo off
pause

plot(loads,'-'), hold on, plot(oloads,'--'), hold off
title('Original loadings (dashed) and estimated loadings (solid)')
xlabel('Variable Number')
ylabel('Loading')