k-means简单实现

k-means简单实现

k-means算法是典型的无监督式聚类算法，目的是将数据自动分为多个组别，如下图数据的分布以及k均值聚类后得到的结果

算法内容

• 选择k个点，作为初始的聚类中心

  k的选择大多情况下认为给定,无法给定时可尝试多个k值分析得到恰当的k值

• 遍历每个数据点,将其与其距离最近的中心点归为一类

• 计算每一类别的均值

• 将均值设为新的中心点

• 再反复上过程

可见该算法为迭代算法

算法实现

• 拿到数据集,先画图观察

def Draw(data):
    data1 = pd.DataFrame(data.get('X'), columns=['X1', 'X2'])
    sb.lmplot('X1', 'X2', data = data1, fit_reg=False)
    plt.show()

• 随机设定初始的中心坐标

def init_center(X, k):
    m, n = X.shape
    centroids = np.zeros((k, n))
    #从数据集中随机选定k个点
    idx = np.random.randint(0, m, k)
    for i in range(k):
        center[i, :] = X[idx[i], :]
    return center

• 将数据集与距离其最近的中心点归为一类

def find_closest_center(X, center):
    m = X.shape[0]
    k = center.shape[0]
    cat = np.zeros(m)

    for i in range(m):
        min_dis = 100000
        for j in range(k):
            dis = np.sum((X[i,:]-center[j,:])**2)
            if dis < min_dis:
                min_dis = dis
                cat[i] = j
    return cat

• 计算每一类的均值

def compute_center(X, idx, k):
    m, n = X.shape
    dis = np.zeros((k, n))
    for i in range(k):
        cnt = np.where(idx == i)
        dis[i,:] = (np.sum(X[cnt,:], axis=1) / len(cnt[0])).ravel()
    return dis

• 更新新的中心点,以上过程迭代进行n次

def run_k_means(X, initial_center, max_iters):
    m,n = X.shape
    k = initial_center.shape[0]
    center = initial_center

    for i in range(max_iters):
        idx = find_closest_center(X, center)
        center = compute_center(X, idx, k)
    return idx, center

完整代码

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sb
from scipy.io import loadmat

def Draw(data):
    data1 = pd.DataFrame(data.get('X'), columns=['X1', 'X2'])
    sb.lmplot('X1', 'X2', data = data1, fit_reg=False)
    plt.show()

def find_closest_center(X, center):
    m = X.shape[0]
    k = center.shape[0]
    cat = np.zeros(m)

    for i in range(m):
        min_dis = 100000
        for j in range(k):
            dis = np.sum((X[i,:]-center[j,:])**2)
            if dis < min_dis:
                min_dis = dis
                cat[i] = j
    return cat

def compute_center(X, idx, k):
    m, n = X.shape
    dis = np.zeros((k, n))
    for i in range(k):
        cnt = np.where(idx == i)
        dis[i,:] = (np.sum(X[cnt,:], axis=1) / len(cnt[0])).ravel()
    return dis

def run_k_means(X, initial_center, max_iters):
    m,n = X.shape
    k = initial_center.shape[0]
    center = initial_center

    for i in range(max_iters):
        idx = find_closest_center(X, center)
        center = compute_center(X, idx, k)
    return idx, center


def init_center(X, k):
    m, n = X.shape
    center = np.zeros((k, n))
    idx = np.random.randint(0, m, k)
    for i in range(k):
        center[i, :] = X[idx[i], :]
    return center

data = loadmat('./data7/ex7data2.mat')
Draw(data)
X = data['X']

center = np.array([[3, 3], [6, 2], [8, 5]])
idx = find_closest_center(X, center)
print(idx[0:3])
dis = compute_center(X, idx, 3)
idx, center = run_k_means(X, center, 10)

cat1 = X[np.where(idx == 0)[0],:]
cat2 = X[np.where(idx == 1)[0],:]
cat3 = X[np.where(idx == 2)[0],:]

fig, ax = plt.subplots(figsize=(12,8))
ax.scatter(cat1[:,0], cat1[:,1], s=30, color='r')
ax.scatter(cat2[:,0], cat2[:,1], s=30, color='g')
ax.scatter(cat3[:,0], cat3[:,1], s=30, color='b')
ax.legend()
plt.show()
print(dis)