pyclustering.cluster.kmeans使用讲解

pyclustering 是一个聚类分析的python库。本文将对其中的kmeans库讲解。

最近本人在用kmeans算法做一些研究，有个想法是把kmeans的距离函数更换，但sklearn并没有提供接口，自己造的轮子效果也并不好。最后找到pyclustering库，因此在这记录一下使用心得。

kmeans训练过程如博客所示。

用到的包

from pyclustering.cluster.center_initializer import kmeans_plusplus_initializer
from pyclustering.cluster.kmeans import kmeans
import numpy as np

1. 初始化形心：

initial_centers = kmeans_plusplus_initializer(x, cluster_num).initialize()

其中x是数据，cluster\_num是簇数目。

2. 实例化kmeans类：

kmeans_instance = kmeans(x, initial_centers, metric=metric)

metric是度量距离，默认是欧式距离，下面细讲。

3. 训练：

kmeans_instance.process()

3. 归类：

clusters = kmeans_instance.get_clusters()

把上面的训练的数据x以列表形式归类。比如数据a,b,c类别分别是1，1，0则返回index列表[[0,0],[1]]

4. 返回形心：

cs = kmeans_instance.get_centers()

5. 预测：

关于预测，这里给出了几种方法以适应不同场景。

• 首先是直接用实例类预测

label = kmeans_instance.predict(x)

• 根据之前得到的clusters

label = np.array([0]*len(x))
for i,sub in enumerate(clusters):
    label[sub] = i

• 根据得到的形心，这里直接封装成函数，metric是度量函数

def Clu_predict(x,cs,class_num,metric = distance_metric(type_metric.EUCLIDEAN)):
    differences = np.zeros((len(x), class_num))
    for index_point in range(len(x)):
        differences[index_point] = [metric(x[index_point], c) for c in cs]
    label = np.argmin(differences, axis=1)
    return label

注意这里效率很满，推荐自己定义矩阵运算。

6. 度量：

• 使用库的度量，以曼哈顿距离为例：

manhattan_metric = distance_metric(type_metric.MANHATTAN)
kmeans_instance = kmeans(x, initial_centers, metric=manhattan_metric)

把type_metric.后面的换掉就行，库提供的距离有

class type_metric(IntEnum):
    """!
 @brief Enumeration of supported metrics in the module for distance calculation between two points.
 """
 ## Euclidean distance, for more information see function 'euclidean_distance'. EUCLIDEAN = 0
 ## Square Euclidean distance, for more information see function 'euclidean_distance_square'.
 EUCLIDEAN_SQUARE = 1
 ## Manhattan distance, for more information see function 'manhattan_distance'.
 MANHATTAN = 2
 ## Chebyshev distance, for more information see function 'chebyshev_distance'.
 CHEBYSHEV = 3
 ## Minkowski distance, for more information see function 'minkowski_distance'.
 MINKOWSKI = 4
 ## Canberra distance, for more information see function 'canberra_distance'.
 CANBERRA = 5
 ## Chi square distance, for more information see function 'chi_square_distance'.
 CHI_SQUARE = 6
 ## Gower distance, for more information see function 'gower_distance'.
 GOWER = 7
 ## User defined function for distance calculation between two points.
 USER_DEFINED = 1000

• 使用自定义距离，以余弦距离为例：

def cosine_distance(a, b):
     a_norm = np.linalg.norm(a)
     b_norm = np.linalg.norm(b)
     similiarity = np.dot(a, b.T)/(a_norm * b_norm)
     dist = 1. - similiarity
     return dist
metric = distance_metric(type_metric.USER_DEFINED, func=cosine_distance)

距离只需实现计算两个点的距离即可。