模糊C均值聚类 python实现

模糊C均值聚类的python实现

存在的问题：

用来判断终止条件的变量dist，是前后两次隶属度矩阵的对应元素做差，取绝对值，再取最大值。用鸢尾花数据集的话，dist的值最低只能到小数点后两位，但是网上别的代码能达到小数点后6、7位。

但是标准互信息指数NMI的值相差不大，都在0.74左右。

import numpy as np
from sklearn import metrics
import time


def read_txt(path, delimiter=','):
    data = np.loadtxt(path, delimiter=delimiter)
    n_row, n_col = data.shape
    n_col = n_col - 1
    y = np.array(data[:, -1], dtype=np.int)
    X = data[:, 0:n_col]

    return X, y



def random_membership_matrix(n_sample, k):
    # 初始化一个隶属度矩阵
    membership_matrix = np.zeros((n_sample, k), dtype=np.float)
    for i in range(n_sample):
        temp = np.random.random(k)  # 生成k个随机浮点数
        temp = temp / sum(temp)
        membership_matrix[i] = temp
    return membership_matrix


def iterate(X, u, k, m):
    # 停止条件来自：https://blog.csdn.net/lyxleft/article/details/88964494
    # 前后两次隶属度矩阵对应元素做差，差的最大值不超过某一阈值
    n_sample, dim = X.shape
    # 先把前一次的隶属度矩阵存起来
    old_u = np.array(u)
    ter = 0.0000001  # 阈值
    iteration = 0
    dist = float('inf')
    max_iter = 20  # 最大迭代次数，这个次数之内达不到阈值，就得重新随机隶属度矩阵了
    center = np.zeros((k, dim), dtype=np.float)  # 类中心，每行存一个类中心

    # 满足最小阈值，或者达到最大迭代次数时停止
    while dist > ter and iteration < max_iter:

        iteration = iteration + 1

        # 根据隶属度矩阵求中心
        for i in range(k):
            up = np.zeros((dim,), dtype=np.float)
            down = 0.0
            for j in range(n_sample):
                up += np.power(u[j, i], m) * X[j]
                down += np.power(u[j, i], m)
            center[i] = up / down

        # 根据中心求隶属度矩阵
        for p in range(n_sample):
            for q in range(k):
                down = 0.0
                up = np.power(np.linalg.norm(X[p] - center[q]), -2/(m-1))
                for e in range(k):
                    down += np.power(np.linalg.norm(X[p] - center[e]), -2/(m-1))
                u[p, q] = up / down


        # 求终止条件
        dist = (abs(u - old_u)).max()

    return center, u


def cluster(membership_matrix):
    # 根据隶属度求聚类结果，把样本点归到隶属度最大的类
    n_sample = membership_matrix.shape[0]
    y_pred = np.zeros((n_sample,), dtype=np.int)
    for i in range(n_sample):
        y_pred[i] = np.argmax(membership_matrix[i])
    return y_pred


if __name__ == '__main__':
    k = 3
    m = 2
    fileName = 'iris'
    X, y = read_txt('./dataset/' + fileName + '.txt')
    n_sample, dim = X.shape
    # 多跑几次取最好结果
    best_y_pred = np.zeros((n_sample,), dtype=np.int)
    best_nmi = 0.0
    center = np.zeros((k, dim), dtype=np.float)
    for d in range(10):
        print('**********第%d次随机隶属度矩阵**********' % d)
        membership_matrix = random_membership_matrix(n_sample, k)
        center, membership_matrix = iterate(X, membership_matrix, k, m)
        y_pred = cluster(membership_matrix)
        nmi = metrics.normalized_mutual_info_score(y, y_pred)
        if nmi > best_nmi:
            best_nmi = nmi
            best_y_pred = y_pred
    print('best_nmi: %f' % best_nmi)
    print(center)