plsrsgcv.m

偏最小二乘算法在MATLAB中的实现

function coeff = plsrsgcv(data,lv,cvit,cvnum)
%PLSRSGCV Generates PLS models for MSPC with cross-validation
%  This function constructs a matrix of PLS models that
%  can be used like a PCA model for multivariate statistical
%  process control (MSPC) purposes. Given a data matrix (data)
%  a PLS model is formed using a maximum of (lv) latent variables
%  that relates each variable to all the others. The actual 
%  number of latent variables used is determined through cross
%  validation using (cvit) test sets with (cvnum) number of
%  samples chosen randomly. The PLS model regression
%  vectors are collected in an output matrix (coeff) which
%  can be used like the I=PP' matrix in PCA.
%  I/O format is: coeff = plsrsgncv(data,lv,cvit,cvnum);
%  
%  Warning:  If you choose to do lots of cross-validations
%  this function can take a long time to execute!

%  Copyright
%  Barry M. Wise
%  1991
%  Modified April 1994

[m,n] = size(data);
if lv >= n
  error('Number of lvs must be <= number of variables - 1')
end
if cvnum >= m
  error('Number of samples in test set must be < total samples')
end
coeff = -eye(n);
disp('  ')
clc
for i = 1:n
  x = [data(:,1:i-1) data(:,i+1:n) data(:,i)];
  press = zeros(lv,1);
  for j = 1:cvit
    home
    s = sprintf('Working on variable number %g, iteration %g',i,j);
    disp(s)
    x = shuffle(x);
    [P,Q,W,T,U,b,ss] = pls(x(1:m-cvnum,1:n-1),x(1:m-cvnum,n),lv);
    mm = conpred(b,W,P,Q,lv);
    cm = cumsum(mm);
    for k = 1:lv
      ypred = x(m-cvnum+1:m,1:n-1)*cm(k,:)';
      press(k,1) = sum((ypred-x(m-cvnum+1:m,n)).^2) + press(k,1);
    end
  end
  plot(press)
  s = sprintf('Cumulative PRESS versus Number of Latent Variables for Variable %g',i); 
  title(s)
  xlabel('Number of Latent Variables')
  ylabel('Cumulative PRESS')
  drawnow
  [a,bb] = min(press);
  s = sprintf('Minimum PRESS for variable %g is at %g',i,bb);
  disp(s)
  home
  s = sprintf('Now forming final PLS model for variable %g',i);
  disp(s)
  [P,Q,W,T,U,b,ss] = pls(x(:,1:n-1),x(:,n),bb);
  mm = conpred(b,W,P,Q,bb);
  cm = (sum(mm))';
  for j = 1:n-1
    if i>j
      coeff(j,i) = cm(j,1);
    end
    if i<=j
      coeff(j+1,i) = cm(j,1);
    end
  end
end
coeff = -1*(coeff);