浅析K-means聚类,具体代码实现
总述
K-means的目的是分类,将一些样本划分成若干蔟,本文将用二维平面上的样本点举例,如图
数学描述
对于蔟 C ( C 1 , C 2 , ⋯ , C n ) C(C_1,C_2,\cdots,Cn) C(C1,C2,⋯,Cn),其质心(可以理解为中心,重心) μ ( μ 1 , μ 2 , ⋯ , μ n ) \mu(\mu_1,\mu_2,\cdots,\mu_n) μ(μ1,μ2,⋯,μn),对于样本点 M ( ( x 1 , y 1 ) , ⋯ ( x n , y n ) ) M((x_1,y_1),\cdots(x_n,y_n)) M((x1,y1),⋯(xn,yn))
质心的计算方法
M
1
−
M
n
M_1-M_n
M1−Mn属于蔟
C
i
C_i
Ci,质心
μ
i
\mu_i
μi,
μ
i
=
1
∣
C
i
∣
∑
M
∈
C
i
M
\mu_i=\frac{1}{|C_i|}\sum_{M\in{C_i}}M
μi=∣Ci∣1M∈Ci∑M
是不是有一点糊涂,我举一个二维平面 的实际例子,有点(1,3)(5,6)它们的质心就为(1+5)/2=3,(3+6)/2=4.5,所以质心为(3,4.5)
选定初始点(质心)
选定k个初始点,通俗讲观察点的分布,直觉判断大概有多少(k)蔟,蔟的中心大概在哪里。
分类样本
如何判断样本是属于哪一蔟?
很简单,样本点距离哪一个蔟的质心越近就属于哪一个蔟,距离
μ
i
\mu_i
μi的距离
D
i
Di
Di
D
i
=
∣
∣
M
i
−
μ
i
∣
∣
D_i = ||M_i-\mu_i||
Di=∣∣Mi−μi∣∣
在二维平面内取L2范数,即使欧几里得距离,点与点之间的距离
(
x
1
−
x
2
)
2
+
(
y
1
−
y
2
)
2
\sqrt{(x_1-x_2)^2+(y_1-y_2)^2}
(x1−x2)2+(y1−y2)2
反复计算质心分类样本
通过选定的初始质心,分类每一个样本点,分类完成后更新每一个蔟的质心,再进行分类,如此直到质心不在发生变化,即是说明每一个点都分类好了。
实现
会用到的两个库
import numpy as np
import matplotlib.pyplot as plt
生成测试数据
可以看到生成了3蔟的数据
s1x = np.random.normal(0, 0.5, 50)
s1y = np.random.normal(0, 0.5, 50)
s2x = np.random.normal(3, 0.5, 50)
s2y = np.random.normal(3, 0.5, 50)
s3x = np.random.normal(3, 0.5, 50)
s3y = np.random.normal(-3, 0.5, 50)
x_t = np.concatenate((s1x, s2x, s3x))
y_t = np.concatenate((s1y, s2y, s3y))
选定初始点(质心)
# 3个蔟,初始化质心
c1 = list()
c2 = list()
c3 = list()
ch1 = (1, 1)
ch2 = (3, 3)
ch3 = (2, -5)
定义函数计算质心
def jszx(c):
sumx = 0
sumy = 0
for k in range(len(c)):
jszx_temp = c[k]
sumx += jszx_temp[0]
sumy += jszx_temp[1]
sumx /= len(c)
sumy /= len(c)
return sumx, sumy
分类样本
for j in range(len(x_t)):
dj1 = (x_t[j] - ch1[0]) ** 2 + (y_t[j] - ch1[1]) ** 2
dj2 = (x_t[j] - ch2[0]) ** 2 + (y_t[j] - ch2[1]) ** 2
dj3 = (x_t[j] - ch3[0]) ** 2 + (y_t[j] - ch3[1]) ** 2
if dj1 == min(dj2, dj3, dj1):
c1.append((x_t[j], y_t[j]))
elif dj2 == min(dj2, dj3, dj1):
c2.append((x_t[j], y_t[j]))
elif dj3 == min(dj2, dj3, dj1):
c3.append((x_t[j], y_t[j]))
反复计算质心分类样本
for i in range(1000):
c1.clear()
c2.clear()
c3.clear()
c1.append(ch1)
c2.append(ch2)
c3.append(ch3)
for j in range(len(x_t)):
dj1 = (x_t[j] - ch1[0]) ** 2 + (y_t[j] - ch1[1]) ** 2
dj2 = (x_t[j] - ch2[0]) ** 2 + (y_t[j] - ch2[1]) ** 2
dj3 = (x_t[j] - ch3[0]) ** 2 + (y_t[j] - ch3[1]) ** 2
if dj1 == min(dj2, dj3, dj1):
c1.append((x_t[j], y_t[j]))
elif dj2 == min(dj2, dj3, dj1):
c2.append((x_t[j], y_t[j]))
elif dj3 == min(dj2, dj3, dj1):
c3.append((x_t[j], y_t[j]))
# 重新计算质心
temp_ch1 = jszx(c1)
temp_ch2 = jszx(c2)
temp_ch3 = jszx(c3)
# 更新质心
if ch1 != temp_ch1 or ch2 != temp_ch2 or ch3 != temp_ch3:
ch1 = temp_ch1
ch2 = temp_ch2
ch3 = temp_ch3
else:
break
整体代码
import numpy as np
import matplotlib.pyplot as plt
# 计算质心
def jszx(c):
sumx = 0
sumy = 0
for k in range(len(c)):
jszx_temp = c[k]
sumx += jszx_temp[0]
sumy += jszx_temp[1]
sumx /= len(c)
sumy /= len(c)
return sumx, sumy
# 随机生成数据
s1x = np.random.normal(0, 0.5, 50)
s1y = np.random.normal(0, 0.5, 50)
s2x = np.random.normal(3, 0.5, 50)
s2y = np.random.normal(3, 0.5, 50)
s3x = np.random.normal(3, 0.5, 50)
s3y = np.random.normal(-3, 0.5, 50)
x_t = np.concatenate((s1x, s2x, s3x))
y_t = np.concatenate((s1y, s2y, s3y))
# plt.plot(x_t, y_t, 'o')
# 3个蔟,初始化质心
c1 = list()
c2 = list()
c3 = list()
ch1 = (1, 1)
ch2 = (3, 3)
ch3 = (2, -5)
# plt.plot(ch1[0], ch1[1], 'ro')
# plt.plot(ch2[0], ch2[1], 'ro')
# plt.show()
# 迭代1000次
for i in range(1000):
c1.clear()
c2.clear()
c3.clear()
c1.append(ch1)
c2.append(ch2)
c3.append(ch3)
for j in range(len(x_t)):
dj1 = (x_t[j] - ch1[0]) ** 2 + (y_t[j] - ch1[1]) ** 2
dj2 = (x_t[j] - ch2[0]) ** 2 + (y_t[j] - ch2[1]) ** 2
dj3 = (x_t[j] - ch3[0]) ** 2 + (y_t[j] - ch3[1]) ** 2
if dj1 == min(dj2, dj3, dj1):
c1.append((x_t[j], y_t[j]))
elif dj2 == min(dj2, dj3, dj1):
c2.append((x_t[j], y_t[j]))
elif dj3 == min(dj2, dj3, dj1):
c3.append((x_t[j], y_t[j]))
# 重新计算质心
temp_ch1 = jszx(c1)
temp_ch2 = jszx(c2)
temp_ch3 = jszx(c3)
# 更新质心
if ch1 != temp_ch1 or ch2 != temp_ch2 or ch3 != temp_ch3:
ch1 = temp_ch1
ch2 = temp_ch2
ch3 = temp_ch3
else:
break
print(i)
# 查看效果
for i in range(len(c1)):
temp = c1[i]
plt.plot(temp[0], temp[1], 'bo')
for i in range(len(c2)):
temp = c2[i]
plt.plot(temp[0], temp[1], 'r*')
for i in range(len(c3)):
temp = c3[i]
plt.plot(temp[0], temp[1], 'g+')
plt.savefig('3class.jpg')
plt.show()