Adaboost

matlab实现了一下《统计机器学习》P140页的例子。adaboost是一个nonUniform的集成算法，提升错误样本的权重，降低正确样本的权重，再用分类器分类训练。

findbest函数，找到最优的decision stump的位置

%% findBest(X,Y,W) return best decision stump
% parameter contains all the information of the basic classification
% parameter:
% ----index--------context-----
%       1        correctRate
%       2      decision stump
%       3         direction
% ----------------
function [parameter,bestResult]=findBest(X,Y,W)

decisionStump=(X+X+1)/2;
bestCorrectRate=0;
bestDecisionStump=0;
bestSignal=0;

for i=1:size(decisionStump,2)
    A=zeros(1,10);
    A(find(X<decisionStump(1,i)))=1;
    A(find(X>=decisionStump(1,i)))=-1;
    correctRate=sum((A==Y).*W);%correct rate
    signal=1;
    result=double(A==Y);
    if correctRate<0.5
        signal=-1;
        correctRate=1-correctRate;
        result=1-double(A==Y);
    end
    if correctRate>bestCorrectRate
        bestCorrectRate=correctRate;
        bestDecisionStump=decisionStump(1,i);
        bestSignal=signal;
        result(find(result==0))=-1;
        bestResult=result;
    end
end

parameter=[bestCorrectRate;bestDecisionStump;bestSignal];

主函数

%% adaboost function 
%
% parameter contains all the information of the basic classification
% parameter:
% ----index--------context-----
%       1        correctRate
%       2      decision stump
%       3         direction
% -----------------------------

% original data
X=[0,1,2,3,4,5,6,7,8,9];
Y=[1,1,1,-1,-1,-1,1,1,1,-1];

%init weight
W=0.1*ones(1,10);
% train model with less error by disition stump
for i=1:4 % max iterator times

    [parameter,bestResult]=findBest(X,Y,W);
    errorRate=1-parameter(1,1);
    amount=1/2*log((1-errorRate)/(errorRate));

    % updata W
    W=W.*exp(-(bestResult.*amount));
    W=W/sum(W,2);

end