Kmean分類

項目鏈接：https://github.com/Wchenguang/gglearn/blob/master/KmeansClassifier/講解/KmeansClassifier.ipynb

1. 首先, 隨機確定 K 個初始點作為質心（不必是數(shù)據(jù)中的點）。
2. 然后將數(shù)據(jù)集中的每個點分配到一個簇中, 具體來講, 就是為每個點找到距其最近的質心, 并將其分配該質心所對應的簇. 這一步完成之后, 每個簇的質心更新為該簇所有點的平均值.
3. 重復上述過程直到數(shù)據(jù)集中的所有點都距離它所對應的質心最近時結束。

            
              import numpy as np
import matplotlib.pyplot as plt

class KmeansClassifier:
    def __init__(self, k, distance_method = "o", random_select = True, 
                    plot = False):
        self.k = k
        self.distance_method = distance_method
        self.random_select = random_select
        self.plot = plot
        if(distance_method == "o"):
            self._dist = self._euler_dist
        else:
            self._dist = self._manhattan_distance
            
    def _euler_dist(self, x1, x):
        return np.sqrt(np.multiply(x-x1, x-x1).sum())
    
    def _manhattan_distance(self, x1, x):
        return np.abs(x-x1).sum()
    
    def _get_nearest(self, x, center_list, dist):
        dists = []
        for center in center_list:
            dists.append(dist(x, center))
        return dists.index(min(dists))
    
    def _fit(self, x, y, dist, x_center_index_list, center_list):
        xy_map = np.hstack((x, x_center_index_list, x_center_index_list))
        for row in xy_map:
            row[-1] = self._get_nearest(row[:-2], center_list, dist)
        flag = np.all(xy_map[:, -1] == xy_map[:, -2])
        return flag, xy_map[:, -1].reshape(-1, 1), center_list
    
    def _random_center_list(self, x, k):
        center_list = np.zeros((x.shape[1], k))
        for col in range(x.shape[1]):
            col_max = np.max(x[:, col])
            col_min = np.min(x[:, col])
            center_list[col, :] = col_min + (col_max - col_min) * np.random.rand(1, k)
        return center_list.T
    
    def _updata_center_list(self, x, x_center_index_list, center_list):
        new_center_list = []
        for index in range(len(center_list)):
            part_x = x[np.where(
                    x_center_index_list[:, -1] == index)]
            if(0 != part_x.size):
                new_center_list.append(np.mean(part_x, axis = 0))
            else:
                new_center_list.append(np.zeros(part_x.shape[1]))
        return new_center_list
    
    def _plot(self, x, x_center_index_list, center_list):
        '''
        數(shù)據(jù)繪制，只能繪制二維
        '''
        center_array = np.array(center_list)
        for index in range(len(center_list)):
            part_x = x[np.where(
                    x_center_index_list[:, -1] == index)]
            plt.scatter(part_x[:, 0], part_x[:, 1])
        plt.scatter(center_array[:, 0], center_array[:, 1], marker = "+")
        plt.show()
    
    def fit(self, x, y, center_list = None):
        if not center_list:
            center_list = self._random_center_list(x, self.k)
        x_center_index_list = np.zeros(x.shape[0]).reshape(-1, 1)
        flag = False
        while(True):
            flag, x_center_index_list, center_list = self._fit(x, 
                                        y, self._dist, x_center_index_list, center_list)
            if(flag):
                break
            center_list = self._updata_center_list(x, x_center_index_list,
                                                      center_list)
            if(self.plot):
                self._plot(x, x_center_index_list, center_list)
        return self

x = np.random.randint(1, 100, (50, 2))
y = np.random.randint(1,4,(10, 1))
kmeans_clf = KmeansClassifier(k = 4, plot=True)
kmeans_clf.fit(x, y)