K-means聚类算法

作用

吧一个数据矩阵分为k堆

原理

首先在数据中随机找K个点, 并且计算其余的点与它们的距离

比如k=3, 数据中有100个点

第一个点离k1近, 那么第一个点就是k1堆的

循环100次, 这样就吧100个点分成了3堆

然后计算这3堆的平均点,用mean函数

这样我们就得到了3个平均点,然后返回上面的步骤,继续计算这3个点与其余100个点的距离,并且分堆

继续计算平均点,继续计算距离,继续分堆,如此反复..

反复几次后,这3个点就不动了,于是数据便被我们分为了3堆

几张图

代码

注意: 如果k值过高返回值可能会出现infinite的情况

function [ centerPoints ] = my_k_mean( data_in, k )
clc;
[data_instance, data_dimension] = size(data_in);

data_run = zeros(data_instance,1);
data_run = [data_in data_run];

if data_instance < k
    error('Data instance less than k');
end
rndPick = randi([1, data_instance - k + 1]);

centerPoints = zeros(k,data_dimension);
for i=1:k
    centerPoints(i,:) = data_in(rndPick + i - 1,:);
end
fClusterChanged = 1;
iteration = 0;

while fClusterChanged == 1
    iteration = iteration + 1;
    disp(['--------------------Iteration: ' num2str(iteration)]); 
    for i=1:data_instance
        pointX = data_in(i,:);
        
        %calcute the distance
        minDist = 0;
        for j=1:k
            dist = pdist([centerPoints(j,:);pointX],'euclidean');
            if minDist > dist || j == 1
                minDist = dist;
                %classifier on the data
                data_run(i,data_dimension + 1) = j;
            end
        end
    end
    
    old_center_points = centerPoints;
    %calcute the new center points
    classes_count = zeros(1,k);
    for i=1:data_instance
        clissfier = data_run(i, data_dimension + 1);
        centerPoints(clissfier,:) = centerPoints(clissfier,:) + data_run(i,1:data_dimension);
        classes_count(clissfier) = classes_count(clissfier) + 1;
    end
    
    for i=1:k
        centerPoints(i,:) = centerPoints(i,:) / classes_count(i);
    end
    
    if or(old_center_points == centerPoints , iteration > 200)
        fClusterChanged = 0;
    else

        fClusterChanged = 1;
    end
end

end