CLRFedu.m

function [EMout,augTrainData,augTestData,trainResults,testResults] = ...
    CLRFedu(Data,covType,depCovType,clampedSetTrain,clampedSetTest,options)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% inputs: Data is the structure containing the following
%         samples => binary training samples. The entries should be 1/2
%         covariates & depCovariates => node and edge covariates in the
%         vertex dependency model
%         covType and depCovType => array indicating whether the covariates
%         are node specific (1), time varying but node constant (2) or both
%         simultaneously (3).
%         clampedSet => array indicating the nodes that prediction is
%         conditioned on. If we want to condition on no nodes we can use an
%         empty matrix.
%         options => structure array giving the input options for structure
%         learning
%
%
% outputs:ParamsEM and sParams => node and edge parameters of the vertex
%         model (saved in EMout)
%         Ecll => expected complete data log likelihood (saved in EMout)
%         aug_covariates and aug_depCovariates are the augmented node and
%         edge covariates of the vertex model extended to include the
%         hidden variables
%         adjmatT => learned structure
%         testResults: structure containing the test results like vertex
%         and edge prediction accuracy as well as co-occurrence statistics.
%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% load data :|
trainSamples = Data.trainSamples;
testSamples = Data.testSamples;
trainCovariates = Data.trainCovariates;
testCovariates = Data.testCovariates;
trainDepCovariates = Data.trainDepCovariates;
testDepCovariates = Data.testDepCovariates;

if nargin < 6
    options.method = 'regCLRG';
    options.diffBias = 0;
end

if ~isfield(options,'diffBias')
    options.diffBias = 0;
end

if isfield(options,'thrsh')
    if ~strcmp(options.method,'CLRG1')
        error('threshold can only be input to CLRG1')
    end
end

if ~isfield(options,'isBox') 
    options.isBox = 0;
end

if ~isfield(options,'isTest') 
    options.isTest = 1;
end


%% Structure Learning
[adjmatT,aug_covariates,aug_depCovariates,optionsEM] =...
    learnStruct(...
    trainSamples,...
    trainCovariates,...
    covType,...
    trainDepCovariates,...
    depCovType,...
    options);

if options.isTest
    [adjmatTtest,aug_covariatesTest,aug_depCovariatesTest,~] =...
        learnStruct(...
        testSamples,...
        testCovariates,...
        covType,...
        testDepCovariates,...
        depCovType,...
        options);
else
    adjmatTtest = [];
    aug_covariatesTest = [];
    aug_depCovariatesTest = [];
end

if options.diffBias
    nSamples = size(trainSamples,2);
    nTot = size(adjmatT,1);
    nTotTest = size(adjmatTtest);
    tmp = repmat(eye(nTot),[1,1,nSamples]);
    tmpTest = repmat(eye(nTotTest),[1,1,nSamples]);
    aug_covariates = [tmp,aug_covariates(:,2:end,:)];
    aug_covariatesTest = [tmpTest,aug_covariatesTest(:,2:end,:)];
end

augTrainData.adjmatT = adjmatT;
augTrainData.aug_covariates = aug_covariates;
augTrainData.aug_depCovariates = aug_depCovariates;

if options.isTest
    augTestData.adjmatT = adjmatTtest;
    augTestData.aug_covariates = aug_covariatesTest;
    augTestData.aug_depCovariates = aug_depCovariatesTest;
else
    augTestData = [];
end

%% Parameter Estimation
optionsEM.stepSize = 1e-2;
[ParamsEM, dParams, ~, Ecll] = ...
    paramsEst(...
    trainSamples,...
    adjmatT,...
    aug_covariates,...
    aug_depCovariates,...
    optionsEM);

EMout.paramsEM = ParamsEM;
EMout.dParams = dParams;
EMout.Ecll = Ecll;
%% Testing
% VP/CP/CA/EA/EAca/EAcp
nSamples = size(trainSamples,2);
clampedCLRFtrain = zeros(length(adjmatT),nSamples);
clampedCLRFtrain(clampedSetTrain,:) = trainSamples(clampedSetTrain,:);

if options.isTest
    clampedCLRFtest = zeros(length(adjmatTtest),nSamples);
    clampedCLRFtest(clampedSetTest,:) = testSamples(clampedSetTest,:);
end

[trainResults.VP,trainResults.CP,trainResults.CA] = ...
    condPredEdu(ParamsEM,dParams,adjmatT,aug_covariates,aug_depCovariates,trainSamples,clampedCLRFtrain,options.isBox);
if options.isTest
    [testResults.VP,testResults.CP,testResults.CA] =...
        condPredEdu(ParamsEM,dParams,adjmatTtest,aug_covariatesTest,aug_depCovariatesTest,testSamples,clampedCLRFtest,options.isBox);
else
    testResults = [];
end

% Co-occurence
[~,~,~,~,co_ocTrain] = CLRFco_occur(ParamsEM,dParams,adjmatT,aug_covariates,aug_depCovariates,trainSamples);
if options.isTest
    [~,~,~,~,co_ocTest] = CLRFco_occur(ParamsEM,dParams,adjmatTtest,aug_covariatesTest,aug_depCovariatesTest,testSamples);
else
    co_ocTest = [];
end

nTrain = size(trainSamples,1);
trainResults.mean_CoOc = zeros(1,nSamples);
for t=1:nSamples
    for i=1:nTrain
        for j=1:i-1
            trainResults.mean_CoOc(t) = trainResults.mean_CoOc(t) + co_ocTrain(i,j,t);
        end
    end
end
trainResults.mean_CoOc = trainResults.mean_CoOc/(nTrain*(nTrain-1)/2);

if options.isTest
    nTest = size(testSamples,1);
    testResults.mean_CoOc = zeros(1,nSamples);
    for t=1:nSamples
        for i=1:nTest
            for j=1:i-1
                testResults.mean_CoOc(t) = testResults.mean_CoOc(t) + co_ocTest(i,j,t);
            end
        end
    end
    testResults.mean_CoOc = testResults.mean_CoOc/(nTest*(nTest-1)/2);
end