tucker.m

强大的多维工具箱.应用在Matlab中,可分析多纬数据结构.直接安装.

function [Factors,G,ExplX,Xm]=tucker(X,Fac,Options,ConstrF,ConstrG,Factors,G);
%TUCKER multi-way tucker model
%
% function [Factors,G,ExplX,Xm]=tucker(X,Fac[,Options[,ConstrF,[ConstrG[,Factors[,G]]]]]);
%
% Change: True LS unimodality now supported.
%
% This algorithm requires access to:
% 'fnipals' 'gsm' 'inituck' 'calcore' 'nmodel' 'nonneg' 'setopts' 'misssum'
% 'missmean' 't3core'
%
% See also:
% 'parafac' 'maxvar3' 'maxdia3' 'maxswd3'
%
% ---------------------------------------------------------           
%             The general N-way Tucker model
% ---------------------------------------------------------
%    
% [Factors,G,ExplX,Xm]=tucker(X,Fac,Options,ConstrF,ConstrG,Factors,G);
% [Factors,G,ExplX,Xm]=tucker(X,Fac);
%
% INPUT
% X        : The multi-way data array.
% Fac      : Row-vector describing the number of factors
%            in each of the N modes. A '-1' (minus one)
%            will tell the algorithm not to estimate factors
%            for this mode, yielding a Tucker2 model.
%            Ex. [3 2 4]
% 
% OPTIONAL INPUT
% Options  : See parafac.
% ConstrF  : Constraints that must apply to 'Factors'.
%            Define a row-vector of size N that describes how
%            each mode should be treated.
%            '0' orthogonality (default)
%            '1' non-negativity
%            '2' unconstrained
%            '4' unimodality and non-negativity.
%            E.g.: [0 2 1] yields ortho in first mode, uncon in the second
%            and non-neg in the third mode.
%            Note: The algorithm uses random values if there are no
%            non-negative components in the iteration intermediates. Thus,
%            if non-negativity is applied, the iterations may be
%            non-monotone in minor sequences.
% ConstrG  : Constraints that must apply to 'G'.
%            '[]' or '0' will not constrain the elements of 'G'.
%            To define what core elements should be allowed, give a core that
%            is 1 (one) on all active positions and zero elsewhere - this boolean
%            core array must have the same dimensions as defined by 'Fac'.
%
% OUTPUT
% Factors  : A row-vector containing the solutions.
% G        : Core array that matches the dimensions defined by 'Fac'.
% ExplX    : Fraction of variation (sums of squares explained)
% Xm       : Xhat (the model of X)
%
% This algorithm applies to the general N-way case, so
% the array X can have any number of dimensions. The
% principles of 'projections' and 'systematic unfolding 
% methodology (SUM)' are used in this algorithm to provide
% a fast approach - also for larger data arrays. This
% algorithm can handle missing values if denoted
% by NaN's. It can also be used to make TUCKER2/1 models by
% properly setting the elements of 'Fac' to -1.
%
% Note: When estimating a Tucker model on data using non-orthogonal factors,
%       the sum of square of the core may differ between models of the
%       same dataset. This is in order since the factors may
%       thus be correlated. However, the expl. var. should always be the same.
%      

% $ Version 2.003 $ Jan 2002 $ Fixed problem with length of factors under special conditions $ CA $ Not compiled $
% $ Version 2.002 $ Jan 2002 $ Fixed reshaping of old input G $ RB $ Not compiled $
% $ Version 2.001 $ July 2001 $ Changed problem with checking if Factors exist (should check if it exists in workspace specifically)$ RB $ Not compiled $
% $ Version 2.00 $ May 2001 $ Changed to array notation $ RB $ Not compiled $
% $ Version 1.12 $ Date 14. Nov. 1999 $ Not compiled $
%
%
% Copyright, 1998 - 
% This M-file and the code in it belongs to the holder of the
% copyrights and is made public under the following constraints:
% It must not be changed or modified and code cannot be added.
% The file must be regarded as read-only. Furthermore, the
% code can not be made part of anything but the 'N-way Toolbox'.
% In case of doubt, contact the holder of the copyrights.
%
% Claus A. Andersson
% Chemometrics Group, Food Technology
% Department of Food and Dairy Science
% Royal Veterinary and Agricultutal University
% Rolighedsvej 30, DK-1958 Frederiksberg, Denmark
% E-mail: [e-mail]
%

DimX = size(X);
X = reshape(X,DimX(1),prod(DimX(2:end)));
FacNew = Fac;
FacNew(find(FacNew==-1)) = DimX(find(FacNew==-1));

format long
format compact
dbg=0;

if nargin==0,
    help('tucker.m');
    error(['Error calling ''tucker.m''. Since no input arguments were given, the ''help'' command was initiated.'])
    return;
end;
if nargin<2,
    help('tucker.m');
    error(['Error calling ''tucker.m''. At least two (2) input arguments must be given. Read the text above.'])
    return;
end;
if size(Fac,2)==1,
    help('tucker.m');
    error(['Error calling ''tucker.m''. ''Fac'' must be a row-vector.'])
end;    

% Initialize system variables
N=size(Fac,2);
Fac_orig=Fac;
finda=find(Fac==-1);
if ~isempty(finda),
    Fac(finda)=zeros(size(finda));
end;
FIdx0=cumsum([1 DimX(1:N-1).*Fac(1:N-1)]);
FIdx1=cumsum([DimX.*Fac]);
pmore=30;
pout=0;
Xm=[];
MissingExist=any(isnan(X(:)));
if MissingExist,
    IdxIsNans=find(isnan(X));
end;
SSX=misssum(misssum(X.^2));

if exist('Options'),
    Options_=Options;
else
    Options_=[0];
end;
load noptiot3.mat;
i=find(Options_);
Options(i)=Options_(i);
if isnan(Options(5)),
    prlvl = 0;
else 
    prlvl = 1;
end;
Options12=Options(1);
Options11=Options12*10;
Options21=Options(2);
Options31=Options(3);
Options41=Options(4);
Options51=Options(5);
Options61=Options(6);
Options71=Options(7);
Options81=Options(8);
Options91=Options(9);
Options101=Options(10);

if ~exist('ConstrF'),
    ConstrF=[];
end;
if isempty(ConstrF),
    ConstrF=zeros(size(DimX));
end;
if ConstrF==0 ,
    ConstrF=zeros(size(DimX));
end;

if ~exist('ConstrG')
    ConstrG=[];
end;
if isempty(ConstrG),
    ConstrG=0;
end;

if exist('Factors')~=1,
    Factors=[];
end;

if ~exist('G'),
    G=[];
else
    G=reshape(G,size(G,1),prod(size(G))/size(G,1));
end;

%Give a status/overview
if prlvl>0,
    fprintf('\n\n');
    fprintf('=================   RESUME  &  PARAMETERS   ===================\n');
    fprintf('Array                 : %i-way array with dimensions (%s)\n',N,int2str(DimX));
    if any(Fac==0),
        fprintf('Model                 : (%s) TUCKER2 model\n',int2str(Fac));
    else
        fprintf('Model                 : (%s) TUCKER3 model\n',int2str(Fac));
    end;   
end

%Mth initialization
txt1=str2mat('derived by SVD (orthogonality constrained).');
txt1=str2mat(txt1,'derived by NIPALS (orthogonality constrained).');
txt1=str2mat(txt1,'derived by Gram-Schmidt (orthogonality constrained).');
txt1=str2mat(txt1,'This mode is not compressed/calculated, i.e., TUCKER2 model.');
txt1=str2mat(txt1,'derived by non-negativity least squares.');
txt1=str2mat(txt1,'derived by unconstrained simple least squares.');
txt1=str2mat(txt1,'unchanged, left as defined in input ''Factors''.');
txt1=str2mat(txt1,'derived by unimodality constrained regression.');
MethodO=1;
for k=1:N,
    UpdateCore(k)=1;
    if ConstrF(k)==0,
        if Fac(k)>0,
            if 0<DimX(k) & DimX(k)<=180,
                Mth(k)=1;
            end;
            if 180<DimX(k) & DimX(k)<=Inf,
                Mth(k)=3;
            end;
            if Fac(k)<=6 & 180<DimX(k),
                Mth(k)=2;
            end;
        end;
        UpdateWithPinv(k)=1; %Update with the L-LS-P-w/Kron approach
        CalcOrdinar(k)=1;
    end;
    if ConstrF(k)==1,
        Mth(k)=5; %nonneg
        MethodO=2; %use the flexible scheme
        UpdateCore(k)=1; %Update the core in this mode
        CalcOrdinar(k)=1;
    end;
    if ConstrF(k)==2,
        Mth(k)=6; %uncon
        MethodO=2; %use the flexible scheme
        UpdateCore(k)=1; %Update the core in this mode
        CalcOrdinar(k)=1;
    end;
    if ConstrF(k)==3,
        Mth(k)=7; %unchanged
        MethodO=2;
        UpdateCore(k)=1; %Update the core in this mode
        CalcOrdinar(k)=1;
    end;
    if ConstrF(k)==4,
        Mth(k)=8; %unimod
        MethodO=2; %use the flexible scheme
        UpdateCore(k)=1; %Update the core in this mode
        CalcOrdinar(k)=1;
    end;   
    if Fac_orig(k)==-1
        Mth(k)=4;
        UpdateCore(k)=0; %Do not update core for this mode
        CalcOrdinar(k)=1;
    end;
    if Options91>=1,
        if prlvl>0,
            if Mth(k)~=4,
                fprintf('Mode %i                : %i factors %s\n',k,Fac(k),txt1(Mth(k),:));
            else
                fprintf('Mode %i                : %s\n',k,txt1(Mth(k),:));
            end;
        end;
    end;
end;

UserFactors=1;
if isempty(Factors),
    UserFactors=0;
else
    ff = [];
    for f=1:length(Factors)
        if ~all(size(Factors{f})==[DimX(f),Fac(f)]), %%
            Factors{f}=rand(DimX(f),Fac(f));%% Added by CA, 27-01-2002
        end;%%
        ff=[ff;Factors{f}(:)];
    end
    OrigFactors=Factors;
    Factors = ff;
end;

usefacinput=0;
if MissingExist,
    if ~UserFactors
        [i j]=find(isnan(X));
        mnx=missmean(X)/3;
        mny=missmean(X')/3;
        n=size(i,1);
        for k=1:n,
            i_=i(k);
            j_=j(k);
            X(i_,j_) = mny(i_) + mnx(j_);
        end;
        mnz=(missmean(mnx)+missmean(mny))/2;
        p=find(isnan(X));
        X(p)=mnz;
    else
        usefacinput=1;
        % Convert to new format
        clear ff,id1 = 0;
        for i = 1:length(DimX) 
            id2 = sum(DimX(1:i).*Fac(1:i));ff{i} = reshape(Factors(id1+1:id2),DimX(i),Fac(i));id1 = id2;
        end
        Fact = ff;
        Xm=nmodel(Fact,reshape(G,Fac_orig));
        Xm = reshape(Xm,DimX(1),prod(DimX(2:end)));
        X(IdxIsNans)=Xm(IdxIsNans);
    end;
    SSMisOld=sum(sum( X(IdxIsNans).^2 ));
    SSMis=SSMisOld;
end;

% Initialize the Factors by some method
UserFactors=1;
if isempty(Factors),
    Factors=inituck(reshape(X,DimX),Fac_orig,2,[]);
    
    % Convert to old factors
    ff = [];
    for f=1:length(Factors)
        ff=[ff;Factors{f}(:)];
    end
    Factors = ff;
    UserFactors=0;
end;

% Initialize the core
Core_uncon=0;
Core_nonneg=0;