• 支持向量机Support Vector Machine做聚类的一个小程序


    对散点进行二分聚类:

    初始聚类中心的选择会影响分类次数甚至是否能成功分类, 算法采用离样本中心很近的两点作为初始聚类点.

    程序如下:

    % 设定分类次数,以自动调整分类精确度
    % part1 得到散点数据并人工指定两个初始聚类中心点
    clear all;clc;close all;  %注意,3c前面不能写东西,会被擦除.
    x = [ [1,4]; [2,3]; [3,4]; [5,3]; [5,1]; [6,3]; [10,3]; [5,5 ] ];
    cluster_count = 4;  % 聚类次数
    len = size(x, 1);
    
    intpx = 0;
    intpy = 0;
    for i = 1:len
        intpx = intpx + x(i, 1);
        intpy = intpy + x(i, 2);
    end
    intpx = intpx/len;
    intpy = intpy/len;
    
    % 注意坐标和矩阵要用方括号而不是圆括号
    theta1 = [intpx + 0.01, intpy + 0.01];
    theta2 = [intpx - 0.01, intpy - 0.01];
    
    figure;
    % part2 循环过程,其中,计数矩阵和索引矩阵的每次初始化都在循环中完成
    for j = 1:cluster_count
        %    初始化索引和计数矩阵
        c = zeros(2, 1);
        %    画聚类中心点, 中点, 求斜率画中垂线
        subplot(2,2,j);
        title(['di',num2str(j)]);
        plot(theta1(1), theta1(2), '*'); hold on;
        plot(theta2(1), theta2(2), '*'); hold on;
        mid1 = (theta1(1) + theta2(1))/2;
        mid2 = (theta1(2) + theta2(2))/2;
        plot(mid1, mid2, '+'); hold on;
        axis([-2 10 -2 10])
        slope = (-1) * (theta1(1) - theta2(1))/(theta1(2) - theta2(2));
        t = 0:0.01:10;
        line = slope*(t-mid1) + mid2;
        plot(t, line); hold on;
        %    判断分类结果, 画出ox区分, 并得到新的双theta
        %    a 分类
        thetanew1 = [0, 0]; thetanew2 = [0, 0];
        for i = 1:len
            if (x(i,1)-theta1(1))^2 + (x(i,2)-theta1(2))^2 < (x(i,1)-theta2(1))^2 + (x(i,2)-theta2(2))^2
                y(i) = 1;
                c(1) = c(1) + 1;
                plot(x(i,1), x(i,2), 'x');hold on;
                thetanew1(1) = thetanew1(1) + x(i,1);
                thetanew1(2) = thetanew1(2) + x(i,2);
            else
                y(i) = 0;
                c(2) = c(2) + 1;
                plot(x(i,1), x(i, 2), 'o');hold on;
                thetanew2(1) = thetanew2(1) + x(i,1);
                thetanew2(2) = thetanew2(2) + x(i,2);
            end
        end
        theta1 = thetanew1/c(1); 
        theta2 = thetanew2/c(2);
        %    b 双theta
        
        axis([-2 10 -2 10])
    end
    
    % 若第一次分类为初始值分类,则可见第三次分类已达最佳
    

     输出图像如下:

    三分聚类:

    % 三分聚类
    % 2c
    clc; close all;
    % clear all is not really necesssary, beccause every variable with the same
    % name
    % 导入要分类的散点数据
    x = [ [1,4]; [2,3]; [3,4]; [5,3]; [5,1]; [6,3]; [10,3]; [5,5 ]; [4, 0]; [3, 0] ];
    %x = [ [1, 1]; [2, 1]; [2,2]; [8,1]; [8,2]; [8,3]; [4, 8]; [5, 8] ];
    cluster_times = 4;
    len = size(x, 1);
    xxall = 0; xyall = 0;
    for i = 1:len
        xxall = xxall + x(i, 1);
        xyall = xyall + x(i, 2);
    end
    
    xysum = [xxall, xyall];
    
    intpx = 0;
    intpy = 0;
    for i = 1:len
        intpx = intpx + x(i, 1);
        intpy = intpy + x(i, 2);
    end
    intpx = intpx/len;
    intpy = intpy/len;
    
    % 注意坐标和矩阵要用方括号而不是圆括号
    %{
    theta1 = [intpx, intpy + 1.01];
    theta2 = [intpx - 1.02, intpy + 1.03];
    theta3 = [intpx + 1.04, intpy - 1.05];
    %}
    
    theta1 = x(1,:);
    theta2 = x(2,:);
    theta3 = x(10,:);
    
    % 判断
    for j = 1:cluster_times
        %    初始化索引和计数矩阵
        c = zeros(3, 1);
        %    画聚类中心点, 中点, 求斜率画中垂线
        %subplot(3,3,j);
        figure;
        % title(['di',num2str(j)]);
    
        % plot(mid1, mid2, '+'); hold on;
        axis([-2 10 -2 10])
        
        
    
        %    判断分类结果, 画出ox区分, 并得到新的双theta
        %    a 分类
        thetanew1 = [0, 0]; thetanew2 = [0, 0]; thetanew3 = [0, 0];
        for i = 1:len
            if (x(i,1)-theta1(1))^2 + (x(i,2)-theta1(2))^2 < (x(i,1)-theta2(1))^2 + (x(i,2)-theta2(2))^2 ...
                    && ((x(i,1)-theta1(1))^2 + (x(i,2)-theta1(2))^2 < (x(i,1)-theta3(1))^2 + (x(i,2)-theta3(2))^2)
                y(i) = 0;
                c(1) = c(1) + 1;
                plot(x(i,1), x(i,2), 'x');hold on;
                thetanew1(1) = thetanew1(1) + x(i,1);
                thetanew1(2) = thetanew1(2) + x(i,2);
            elseif (x(i,1)-theta2(1))^2 + (x(i,2)-theta2(2))^2 < (x(i,1)-theta1(1))^2 + (x(i,2)-theta1(2))^2 ...
                    && ((x(i,1)-theta2(1))^2 + (x(i,2)-theta2(2))^2 < (x(i,1)-theta3(1))^2 + (x(i,2)-theta3(2))^2)
                y(i) = 1;
                c(2) = c(2) + 1;
                plot(x(i,1), x(i, 2), 'o'); hold on;
                thetanew2(1) = thetanew2(1) + x(i,1);
                thetanew2(2) = thetanew2(2) + x(i,2);
            else
                y(i) = 2;
                c(3) = c(3) + 1;
                plot(x(i, 1), x(i, 2), '+'); hold on;
                thetanew3(1) = thetanew3(1) + x(i, 1);
                thetanew3(2) = thetanew3(2) + x(i, 2);
            end
        end
        theta1 = thetanew1/c(1);
        theta2 = thetanew2/c(2);
        theta3 = thetanew3/c(3);
        
        mid12 = (theta1 + theta2)/2;
        mid23 = (theta2 + theta3)/2;
        mid31 = (theta3 + theta1)/2;
        slope12 = (-1) * (theta1(1) - theta2(1))/(theta1(2) - theta2(2));  %负倒数通过交换分子分母得到
        slope23 = (-1) * (theta2(1) - theta3(1))/(theta2(2) - theta3(2));
        slope31 = (-1) * (theta3(1) - theta1(1))/(theta3(2) - theta1(2));
        
        t = 0:0.01:10;
        %t = 4.2:0.01:10;
        line12 = slope12*(t-mid12(1)) + mid12(2);
        plot(t, line12); hold on;
        
        %t = 0:0.01:4.2;
        line23 = slope23*(t-mid23(1)) + mid23(2);
        plot(t, line23); hold on;
        
        %t = 4.2:0.01:10;
        line31 = slope31*(t-mid31(1)) + mid31(2);
        plot(t, line31); hold on;
        %plot(theta1(1), theta1(2), '*'); 
        %plot(theta2(1), theta2(2), '*'); 
        %plot(theta3(1), theta3(2), '*');
        % 运行结果不能达到预期时先不要否定算法而是先检查一下细节
        %    b 双theta
        
        axis([-2 10 -2 10])
    end
    

     输出如下:

    初始点theta1, theta2, theta3 选x1, x2, x10:

    初始点选x1, x2, x7 就会陷入局部最优:

  • 相关阅读:
    人机猜拳
    String字符串与字符(char类型)数组互相转换
    String字符串与字符(char类型)数组互相转换
    pytorch(五)超参调试过程的tensorboard可视化
    pytorch(三)建立模型(网络)
    pytorch(二)数据准备工作ETL
    pytorch(一)张量基础及通用操作
    中东冲突及圣城耶路撒冷(希伯来语为和平之城)
    五线谱基础及弹奏姿势
    相机主距与焦距
  • 原文地址:https://www.cnblogs.com/zhangziyan/p/9571733.html
Copyright © 2020-2023  润新知