cce.m

CCE is a multi-instance learning method solving multi-instance problems through adapting multi-insta

function output=CCE(train_bags,test_bags,traintarget,num,gamma,cost)
%CCE implements the algorithm described in [1], where Gaussian kernel LibSVM [2] is used as the base learner for transformed feature vectors.
%    Syntax
%
%      Outputs=CCE(train_bags,test_bags,traintarget,num,gamma,cost)
%
%    Description
%
%       CCE takes,
%           train_bags        - An Mx1 cell, where the jth instance of the ith training bag is stored in train_bags{i,1}(j,:)
%           test_bags            - An Nx1 cell, where the jth instance of the ith test bag is stored in test_bags{i,1}(j,:)
%           traintarget         - A 1xM array, if the ith training bag is positive then traintarget(1,i)=+1, otherwise equals -1
%           num                  - An 1xD array, where the required number of clusters for the ith individual classifier in the ensemble is sotred in
%                                  num(1,i)
%           gamma                - The gamma parameter for the gaussian kernel svm, default=1
%           cost                 - The C-parameter for the gaussian kernel svm, default=1
%      and returns,
%           output              - A 1xN array, where the output of the ith test bag is stored in output(1,i), +1 for positive and -1 for negative
%
% [1]Z.-H. Zhou and M.-L. Zhang. Solving multi-instance problems with classifier ensemble based on constructive clustering. Knowledge and Information
%    Systems, 2007, 11(2): 155-170.
% [2]C.-C. Chang and C.-J. Lin. LIBSVM: a library for support vector machines. Technical Report, Department of Computer Science and Information
%    Engineering, National Taiwan University, Taipei, China, 2001.

    if(nargin<4)
        error('Not enough input parameters, please check again.');
    end

    if(nargin<6)
        cost=1;
    end
    if(nargin<5)
        gamma=1;
    end

    [num_train,tempvalue]=size(train_bags);
    [num_test,tempvalue]=size(test_bags);
    [tempvalue,en_size]=size(num);
    
    Outputs=zeros(num_test,en_size);
    
    instances=[];
    for i=1:num_train
        instances=[instances;train_bags{i,1}];
    end
    
    for iter=1:en_size
        disp(strcat('Building the ',num2str(iter),'-th individual classifier'));
        centers=k_means(instances,num(1,iter));
        
        trainset=zeros(num(1,iter),num_train);
        testset=zeros(num(1,iter),num_test);
        
        for i=1:num_train
            tempbag=train_bags{i,1};
            [tempsize,tempvalue]=size(tempbag);
            for j=1:tempsize
                tempdist=zeros(1,num(1,iter));
                for k=1:num(1,iter)
                    tempdist(1,k)=(tempbag(j,:)-centers(1,k))*(tempbag(j,:)-centers(1,k))';
                end
                [tempmin,index]=min(tempdist);
                trainset(index,i)=trainset(index,i)+1;
            end            
        end
        trainset=(trainset>=1);
        
        for i=1:num_test
            tempbag=test_bags{i,1};
            [tempsize,tempvalue]=size(tempbag);
            for j=1:tempsize
                tempdist=zeros(1,num(1,iter));
                for k=1:num(1,iter)
                    tempdist(1,k)=(tempbag(j,:)-centers(1,k))*(tempbag(j,:)-centers(1,k))';
                end
                [tempmin,index]=min(tempdist);
                testset(index,i)=testset(index,i)+1;
            end            
        end 
        testset=(testset>=1);
        
        [AlphaY, SVs, Bias, Parameters, nSV, nLabel] = RbfSVC(trainset,traintarget,gamma,cost);
        Labels=ones(1,num_test);
        [ClassRate, DecisionValue, Ns, ConfMatrix, PreLabels]= SVMTest(testset, Labels, AlphaY, SVs, Bias,Parameters, nSV, nLabel);
        Outputs(:,iter)=PreLabels';
    end
    
    output=((sum(Outputs')>=0)-0.5)*2;