高斯混合模型

高斯混合模型一维应用

一、代码

import numpy as np
import matplotlib.pyplot as plt 
from math import e

pi = 3.1415926535

def Gaussian_distribution(u,v,x):  #正态分布密度函数
    y = 1 / (v * pow(2*pi, 0.5))*np.exp(-((x-u)**2)/(2*v**2))  # 概率密度函数公式
    return y

u1 = 1
u2 = 2
x1 = u1 + np.random.uniform(0,1,100) #random()产生的随机数符合正态分布
x2 = u2 + np.random.uniform(0,1,100)
x = np.zeros([2,100])
x[0],x[1] = x1,x2
x = x.reshape(1,200)
x = x.ravel()

x1 = np.array(x1) 
x2 = np.array(x2)
x1_mean = np.mean(x1)  #以产生随机数的均值作为正态分布的均值
x2_mean = np.mean(x2)
u1 = x1_mean
u2 = x2_mean    
C1 = 1.0/100*(x1-x1_mean).T@(x1-x1_mean)#以产生随机数的方差作为正态分布的方差
C2 = 1.0/100*(x2-x2_mean).T@(x2-x2_mean)
v1 = np.sqrt(C1)
v2 = np.sqrt(C2)

k1 = 0.5
k2 = 0.5
m1 = x1_mean #数据初始化
m2 = x2_mean
n1 = v1
n2 = v2
maxtime = 100 #最大迭代次数


for i in range(0,maxtime):
    q1_sum = 0
    q2_sum = 0
    q1_xsum = 0
    q2_xsum = 0
    q1 = 0
    q2 = 0
    s1 = 0 
    s2 = 0
    for j in range(200): #EM算法
        a1 = k1*Gaussian_distribution(m1,n1,x[j])
        a2 = k2*Gaussian_distribution(m2,n2,x[j])
        q1 = a1/(a1+a2)
        q2 = a2/(a1+a2)
        q1_sum = q1_sum+q1
        q2_sum = q2_sum+q2
        q1_xsum  = q1_xsum+q1*x[j]
        q2_xsum  = q2_xsum+q2*x[j]
#    print(q1_xsum,q1_sum,q1_xsum/q1_sum,q2_xsum,q2_sum,q2_xsum/q2_sum)
    mm1 = q1_xsum/q1_sum
    mm2 = q2_xsum/q2_sum
    for j in range(200):
        a1 = k1*Gaussian_distribution(m1,n1,x[j])
        a2 = k2*Gaussian_distribution(m2,n2,x[j])
        q1 = a1/(a1+a2)
        q2 = a2/(a1+a2)
        s1 = s1+q1*(x[j]-mm1)*(x[j]-mm1)
        s2 = s2+q2*(x[j]-mm2)*(x[j]-mm2)
    m1 = mm1 #状态更新
    m2 = mm2
    n1 = s1/q1_sum
    n2 = s2/q2_sum
    if(((k1-q1_sum/200)/k1<0.01) &((k2-q2_sum/200)/k2 < 0.01)):  #迭代误差小于1%,退出迭代
        break
    k1 = q1_sum/200
    k2 = q2_sum/200


sum = 0
for j in range(0,100):
    a1 = k1*Gaussian_distribution(m1,n1,x[j])
    a2 = k2*Gaussian_distribution(m2,n2,x[j])
    q1 = a1/(a1+a2)
    q2 = a2/(a1+a2)
    if q1<q2:
        plt.scatter(x[j],a1,marker = '*',color = 'b')
    else:
        plt.scatter(x[j],a1,marker = '.',color = 'b')
    if q1 < q2:
        sum = sum+1
print(sum/100)

sum = 0 
for j in range(100,200):
    a1 = k1*Gaussian_distribution(m1,n1,x[j])
    a2 = k2*Gaussian_distribution(m2,n2,x[j])
    # if a2 >1:
    #     print("warning",m2,v2,x[j])
    q1 = a1/(a1+a2)
    q2 = a2/(a1+a2)
    if q1<q2:
        plt.scatter(x[j],a2,marker = '*',color = 'r')
    else:
        plt.scatter(x[j],a2,marker = '.',color = 'r')
    if q1 > q2:
        sum = sum+1
print(sum/100)

print(m1,m2,n1,n2)
print(u1,u2,v1,v2)
print(k1,k2)
plt.show()
    
    
    

二、原理

高斯混合模型及EM算法

三、结果分析

蓝色的点和红色的点分别表示原有数据原本的正态分布，圆点和星号表示数据被分到了哪个正态分布。

四、存在问题

由于算法问题，导致如果数据中，两个正态分布十分接近，会导致只有一个正态分布。如果是随机产生的数据，重复多次运行可能能解决问题。