K-means 算法是一种简单有效的无监督学习方法,它可以有效地将多维空间(用N表示)中的点聚成一个个紧密的簇。
K-means算法的优化目标是使求出K个中心点,使每一个点到该点的欧氏距离平方之和尽量小。
简单来说就是把一个分到一个类中的所有数据点的每一维相加,得一个向量。然后,该向量的每一维除以该类的点的个数。这样得的向量就是该类的中心(centroid).
算法的思路如下:
1. 初始化K个中心点。
这K个点可以是在所有输入数据点中随机抽取的,也可以是取前K个点,也可以是从N维空间中任意一个点。初始聚类中心点的选择,对最终聚类结果有一定影响;
2. 对任意一个数据点,求与它最近的中心点,并认为该数据点属于该中心点所代表的类。对于M(假设共有M个数据点)个数据点,分别计算每个点与K个当前的中心点的欧氏距离平方值,点x_i与哪个中心点(如c_j)的欧氏距离平方最小那么它就分成该类。(该过程可以求出一些指标,用于终止程序。如,求出整体欧氏距离之和);
3. 更新每个类的中心点。
4. 由 2 得出的指标判断是否可以终止:否,进行 2 ;是,终止,并给出中心点信息。
#include <iostream>
#include <fstream>
#include <cmath>
#include <vector>
#include <opencv2\opencv.hpp>
using namespace cv;
using namespace std;
#define k 3
cv::Scalar colorBar[]=
{
{0, 0, 255},
{0, 255, 0},
{255, 0, 0},
{ 0, 0, 0 },
{ 0, 255, 255 },
{ 255, 0, 255 }
};
//存放元组的属性信息
struct Tuple {
float attr1;
float attr2;
};
//计算两个元组间的欧几里距离
float getDistXY(Tuple t1, Tuple t2)
{
return sqrt((t1.attr1 - t2.attr1) * (t1.attr1 - t2.attr1) + (t1.attr2 - t2.attr2) * (t1.attr2 - t2.attr2));
}
//根据质心,决定当前元组属于哪个簇
int clusterOfTuple(Tuple means[], Tuple tuple) {
float dist = getDistXY(means[0], tuple);
float tmp;
int label = 0;//标示属于哪一个簇
for (int i = 1; i<k; i++) {
tmp = getDistXY(means[i], tuple);
if (tmp<dist) { dist = tmp; label = i; }
}
return label;
}
//获得给定簇集的平方误差
float getVar(vector<Tuple> clusters[], Tuple means[]) {
float var = 0;
for (int i = 0; i < k; i++)
{
vector<Tuple> t = clusters[i];
for (int j = 0; j< t.size(); j++)
{
var += getDistXY(t[j], means[i]);
}
}
//cout<<"sum:"<<sum<<endl;
return var;
}
//获得当前簇的均值(质心)
Tuple getMeans(vector<Tuple> cluster) {
int num = cluster.size();
double meansX = 0, meansY = 0;
Tuple t;
for (int i = 0; i < num; i++)
{
meansX += cluster[i].attr1;
meansY += cluster[i].attr2;
}
t.attr1 = meansX / num;
t.attr2 = meansY / num;
return t;
//cout<<"sum:"<<sum<<endl;
}
void KMeans(vector<Tuple> tuples) {
vector<Tuple> clusters[k];
Tuple means[k];
int i = 0;
//默认一开始将前K个元组的值作为k个簇的质心(均值)
for (i = 0; i<k; i++) {
means[i].attr1 = tuples[i].attr1;
means[i].attr2 = tuples[i].attr2;
}
int lable = 0;
//根据默认的质心给簇赋值
for (i = 0; i != tuples.size(); ++i) {
lable = clusterOfTuple(means, tuples[i]);
clusters[lable].push_back(tuples[i]);
}
cv::Mat showResult = cv::Mat::zeros(10, 10, CV_8UC3);
showResult.setTo(255);
//输出刚开始的簇
for (lable = 0; lable<k; lable++) {
cout << "第" << lable + 1 << "个簇:" << endl;
vector<Tuple> t = clusters[lable];
for (i = 0; i< t.size(); i++)
{
cout << "(" << t[i].attr1 << "," << t[i].attr2 << ")" << " ";
showResult.at<Vec3b>(t[i].attr1, t[i].attr2)[0] = colorBar[lable].val[0];
showResult.at<Vec3b>(t[i].attr1, t[i].attr2)[1] = colorBar[lable].val[1];
showResult.at<Vec3b>(t[i].attr1, t[i].attr2)[2] = colorBar[lable].val[2];
}
cout << endl;
}
float oldVar = -1;
float newVar = getVar(clusters, means);
while (abs(newVar - oldVar) >= 0) //当新旧准则函数值不发生明显变化时,算法终止
{
for (i = 0; i < k; i++) //更新每个簇的中心点
{
means[i] = getMeans(clusters[i]);
//cout<<"means["<<i<<"]:"<<means[i].attr1<<" "<<means[i].attr2<<endl;
}
oldVar = newVar;
newVar = getVar(clusters, means); //计算新的准则函数值
for (i = 0; i < k; i++) //清空每个簇
{
clusters[i].clear();
}
//根据新的质心获得新的簇
for (i = 0; i != tuples.size(); ++i) {
lable = clusterOfTuple(means, tuples[i]);
clusters[lable].push_back(tuples[i]);
}
//输出当前的簇
showResult.setTo(255);
for (lable = 0; lable<k; lable++) {
cout << "第" << lable + 1 << "个簇:" << endl;
vector<Tuple> t = clusters[lable];
for (i = 0; i< t.size(); i++)
{
cout << "(" << t[i].attr1 << "," << t[i].attr2 << ")" << " ";
showResult.at<Vec3b>(t[i].attr1, t[i].attr2)[0] = colorBar[lable].val[0];
showResult.at<Vec3b>(t[i].attr1, t[i].attr2)[1] = colorBar[lable].val[1];
showResult.at<Vec3b>(t[i].attr1, t[i].attr2)[2] = colorBar[lable].val[2];
}
cout << endl;
}
}
}
int main() {
ifstream infile;
infile.open("..\\data\\in.txt", ios::in);
if (!infile) {
cout << "不能打开输入的文件" << endl;
return 0;
}
int count = 0;
vector<Tuple> tuples;
Tuple tuple;
//从文件流中读入数据
while (!infile.eof()) {
count++;
if (count % 2 == 1) infile >> tuple.attr1;
else {
infile >> tuple.attr2;
tuples.push_back(tuple);
}
}
//输出文件中的元组信息
for (vector<Tuple>::size_type ix = 0; ix != tuples.size(); ++ix)
cout << "(" << tuples[ix].attr1 << "," << tuples[ix].attr2 << ")" << " ";
cout << endl;
KMeans(tuples);
return 0;
}
结果如图: