机器学习技法第二次作业编程题1

用adaboost的方法对所给数据点进行分类。

matlab代码：

data_train = load('hw2_adaboost_train.dat');
figure(1);
plot(data_train(data_train(:,3)==1,1),data_train(data_train(:,3)==1,2),'r*'); hold on
plot(data_train(data_train(:,3)==-1,1),data_train(data_train(:,3)==-1,2),'bo'); %画图


T = 300; 
N = size(data_train,1);
feature = size(data_train,2) - 1;
%for iter=1:T  %伪代码逻辑示意
%   for i=1:2  
%       for j=1:N
%           thred遍历
%       end 
%       最优（阈值，s）
%   end
%   最优（阈值，s，i）
%   更新u
%end
for i = 1:feature
    %先对feature进行排序，会比较节约时间，三维数组存放
    sdata_f(:,:,i) = sort(data_train,i,'ascend');
end
q16 = [];
theta_iter = ones(T,1); %记录每次迭代获得的参数
s_iter = ones(T,1);
i_iter = ones(T,1);
alpha_iter = ones(T,1);
u = ones(N,1)/N;  %u权值初始化,原数据，没有排序前的顺序
in_pre = zeros(N,1);
for iter = 1:T
    err_iter = 1;
    for i = 1:2
        sdata = sdata_f(:,:,i);
        %初始化，theta负无穷
        theta = -inf;  
        pre = sign(data_train(:,i)-theta); %先默认s=1,之后纠正
        errj = sum(u(pre~=data_train(:,3)))/sum(u); 
        sj = 1;
        if errj > 0.5
            sj = -1;
            errj = 1-errj;
        end
        for j = 1:N-1
            s = 1;
            theta_temp = (sdata(j,i)+sdata(j+1,i))/2; 
            pre = sign(data_train(:,i)-theta_temp);
            err_temp = sum(u(pre~=data_train(:,3)))/sum(u); 
            if err_temp > 0.5
                s = -1;
                err_temp = 1-err_temp;
            end
            if err_temp < errj
                errj = err_temp; %更新errj
                theta = theta_temp;  %记录theta，s
                sj = s;
            end
        end
        if errj < err_iter
            err_iter = errj;   
            theta_iter(iter) = theta; %迭代记录
            s_iter(iter) = sj;
            i_iter(iter) = i;
        end
    end
    q16 = [q16 err_iter]; %question 16
    diamonds_t = sqrt((1-err_iter)/err_iter);
    alpha_iter(iter) = log(diamonds_t);
    pre = s_iter(iter)*sign(data_train(:,i_iter(iter))-theta_iter(iter));
    in_pre = alpha_iter(iter).* pre + in_pre;
    u(pre~=data_train(:,3)) = u(pre~=data_train(:,3)) * diamonds_t; %update u.
    u(pre==data_train(:,3)) = u(pre==data_train(:,3)) / diamonds_t;
    sum(u); %question 15
    figure(2)
    hold on
    if i_iter(iter)==1
        plot([theta_iter(iter),theta_iter(iter)],[0,1],'color',[0.8,0.8,0.8]);
    else
        plot([0,1],[theta_iter(iter),theta_iter(iter)],'color',[0.8,0.8,0.8]);
    end
end
% in_pre = sign(in_pre);
% plot(data_train(in_pre==1,1),data_train(in_pre==1,2),'r*'); %画图
% hold on
% plot(data_train(in_pre==-1,1),data_train(in_pre==-1,2),'bo');

%test
data_test = load('hw2_adaboost_test.dat');
plot(data_test(data_test(:,3)==1,1),data_test(data_test(:,3)==1,2),'r*');hold on %画图
plot(data_test(data_test(:,3)==-1,1),data_test(data_test(:,3)==-1,2),'bo'); %画图
out_pre = zeros(size(data_test,1),1);
for i=1:T
    out_pre = s_iter(i).*sign(data_test(:,i_iter(i))-theta_iter(i)) + out_pre;
end
out_pre = sign(out_pre);
figure(3)
plot(data_test(out_pre==1,1),data_test(out_pre==1,2),'r*'); hold on %画图
plot(data_test(out_pre==-1,1),data_test(out_pre==-1,2),'bo');
sum(out_pre~=data_test(:,3))/size(data_test,1)

代码有点乱，不过逻辑大概如此。