我从头开始使用 K++ 初始化(使用固定种子和 5 个质心。将它添加到您想要的质心数量应该不会太难):
# K++ initialization Algorithm:
import random
def initialize(X, K):
C = [X[0]]
for k in range(1, K):
D2 = scipy.array([min([scipy.inner(c-x,c-x) for c in C]) for x in X])
probs = D2/D2.sum()
cumprobs = probs.cumsum()
np.random.seed(20) # fixxing seeds
#random.seed(0) # fixxing seeds
r = scipy.rand()
for j,p in enumerate(cumprobs):
if r < p:
i = j
break
C.append(X[i])
return C
a = initialize(data2,5) # "a" is the centroids initial array... I used 5 centroids
# Now the Fuzzy c means algorithm:
m = 1.5 # Fuzzy parameter (it can be tuned)
r = (2/(m-1))
# Initial centroids:
c1,c2,c3,c4,c5 = a[0],a[1],a[2],a[3],a[4]
# prepare empty lists to add the final centroids:
cc1,cc2,cc3,cc4,cc5 = [],[],[],[],[]
n_iterations = 10000
for j in range(n_iterations):
u1,u2,u3,u4,u5 = [],[],[],[],[]
for i in range(len(data2)):
# Distances (of every point to each centroid):
a = LA.norm(data2[i]-c1)
b = LA.norm(data2[i]-c2)
c = LA.norm(data2[i]-c3)
d = LA.norm(data2[i]-c4)
e = LA.norm(data2[i]-c5)
# Pertenence matrix vectors:
U1 = 1/(1 + (a/b)**r + (a/c)**r + (a/d)**r + (a/e)**r)
U2 = 1/((b/a)**r + 1 + (b/c)**r + (b/d)**r + (b/e)**r)
U3 = 1/((c/a)**r + (c/b)**r + 1 + (c/d)**r + (c/e)**r)
U4 = 1/((d/a)**r + (d/b)**r + (d/c)**r + 1 + (d/e)**r)
U5 = 1/((e/a)**r + (e/b)**r + (e/c)**r + (e/d)**r + 1)
# We will get an array of n row points x K centroids, with their degree of pertenence
u1.append(U1)
u2.append(U2)
u3.append(U3)
u4.append(U4)
u5.append(U5)
# now we calculate new centers:
c1 = (np.array(u1)**2).dot(data2) / np.sum(np.array(u1)**2)
c2 = (np.array(u2)**2).dot(data2) / np.sum(np.array(u2)**2)
c3 = (np.array(u3)**2).dot(data2) / np.sum(np.array(u3)**2)
c4 = (np.array(u4)**2).dot(data2) / np.sum(np.array(u4)**2)
c5 = (np.array(u5)**2).dot(data2) / np.sum(np.array(u5)**2)
cc1.append(c1)
cc2.append(c2)
cc3.append(c3)
cc4.append(c4)
cc5.append(c5)
if (j>5):
change_rate1 = np.sum(3*cc1[j] - cc1[j-1] - cc1[j-2] - cc1[j-3])/3
change_rate2 = np.sum(3*cc2[j] - cc2[j-1] - cc2[j-2] - cc2[j-3])/3
change_rate3 = np.sum(3*cc3[j] - cc3[j-1] - cc3[j-2] - cc3[j-3])/3
change_rate4 = np.sum(3*cc4[j] - cc4[j-1] - cc4[j-2] - cc4[j-3])/3
change_rate5 = np.sum(3*cc5[j] - cc5[j-1] - cc5[j-2] - cc5[j-3])/3
change_rate = np.array([change_rate1,change_rate2,change_rate3,change_rate4,change_rate5])
changed = np.sum(change_rate>0.0000001)
if changed == 0:
break
print(c1) # to check a centroid coordinates c1 - c5 ... they are the last centroids calculated, so supposedly they converged.
print(U) # this is the degree of pertenence to each centroid (so n row points x K centroids columns).
我知道它不是很 Python,但我希望它可以成为您完整的模糊 C 均值算法的起点。我认为“软聚类”是数据不易分离的方式(例如,当“t-SNE 可视化”将所有数据显示在一起而不是清楚地分开显示组时。在这种情况下,强制数据严格属于只有一个集群可能是危险的)。我会尝试使用 m = 1.1 到 m = 2.0,这样您就可以看到模糊参数如何影响 pertenence 矩阵。