性能度量

由于聚类算法不依赖于样本的真实类标，就不能像监督学习的分类那般，通过计算分对分错（即精确度或错误率）来评价学习器的好坏或作为学习过程中的优化目标。一般聚类有两类性能度量指标：外部指标和内部指标。

外部指标

即将聚类结果与某个参考模型的结果进行比较，以参考模型的输出作为标准，来评价聚类好坏。假设聚类给出的结果为λ，参考模型给出的结果是λ*，则我们将样本进行两两配对，定义：

%7C%5Clambda_i%3D%5Clambda_j%2C%5Clambda_i%5E%20%3D%5Clambda_j%5E%2Ci%3Cj%20%5C%7D%20%EF%BC%8C%E5%8F%82%E8%80%83%E7%BB%93%E6%9E%9C%E5%90%8C%E7%B1%BB%E7%B0%87%EF%BC%8C%E8%81%9A%E7%B1%BB%E7%BB%93%E6%9E%9C%E5%90%8C%E7%B1%BB%E7%B0%87%20%20%0A#card=math&code=a%20%3D%20%7CSS%7C%2C%20SS%3D%5C%7B%28x_i%2Cx_j%29%7C%5Clambda_i%3D%5Clambda_j%2C%5Clambda_i%5E%2A%20%3D%5Clambda_j%5E%2A%2Ci%3Cj%20%5C%7D%20%EF%BC%8C%E5%8F%82%E8%80%83%E7%BB%93%E6%9E%9C%E5%90%8C%E7%B1%BB%E7%B0%87%EF%BC%8C%E8%81%9A%E7%B1%BB%E7%BB%93%E6%9E%9C%E5%90%8C%E7%B1%BB%E7%B0%87%20%20%0A&id=by1aB)

%7C%5Clambda_i%3D%5Clambda_j%2C%5Clambda_i%5E%20%5Cneq%20%5Clambda_j%5E%2Ci%3Cj%20%5C%7D%20%EF%BC%8C%E5%8F%82%E8%80%83%E7%BB%93%E6%9E%9C%E4%B8%8D%E5%90%8C%E7%B1%BB%E7%B0%87%EF%BC%8C%E8%81%9A%E7%B1%BB%E7%BB%93%E6%9E%9C%E5%90%8C%E7%B1%BB%E7%B0%87%20%20%0A#card=math&code=b%20%3D%20%7CSD%7C%2C%20SD%3D%5C%7B%28x_i%2Cx_j%29%7C%5Clambda_i%3D%5Clambda_j%2C%5Clambda_i%5E%2A%20%5Cneq%20%5Clambda_j%5E%2A%2Ci%3Cj%20%5C%7D%20%EF%BC%8C%E5%8F%82%E8%80%83%E7%BB%93%E6%9E%9C%E4%B8%8D%E5%90%8C%E7%B1%BB%E7%B0%87%EF%BC%8C%E8%81%9A%E7%B1%BB%E7%BB%93%E6%9E%9C%E5%90%8C%E7%B1%BB%E7%B0%87%20%20%0A&id=cZkKe)

%7C%5Clambda_i%20%5Cneq%20%5Clambda_j%2C%5Clambda_i%5E%20%3D%5Clambda_j%5E%2Ci%3Cj%20%5C%7D%20%EF%BC%8C%E5%8F%82%E8%80%83%E7%BB%93%E6%9E%9C%E5%90%8C%E7%B1%BB%E7%B0%87%EF%BC%8C%E8%81%9A%E7%B1%BB%E7%BB%93%E6%9E%9C%E4%B8%8D%E5%90%8C%E7%B1%BB%E7%B0%87%20%20%0A#card=math&code=c%20%3D%20%7CDS%7C%2C%20DS%3D%5C%7B%28x_i%2Cx_j%29%7C%5Clambda_i%20%5Cneq%20%5Clambda_j%2C%5Clambda_i%5E%2A%20%3D%5Clambda_j%5E%2A%2Ci%3Cj%20%5C%7D%20%EF%BC%8C%E5%8F%82%E8%80%83%E7%BB%93%E6%9E%9C%E5%90%8C%E7%B1%BB%E7%B0%87%EF%BC%8C%E8%81%9A%E7%B1%BB%E7%BB%93%E6%9E%9C%E4%B8%8D%E5%90%8C%E7%B1%BB%E7%B0%87%20%20%0A&id=zv5oI)

%7C%5Clambda_i%20%5Cneq%20%5Clambda_j%2C%5Clambda_i%5E%20%5Cneq%20%5Clambda_j%5E%2Ci%3Cj%20%5C%7D%20%EF%BC%8C%E5%8F%82%E8%80%83%E7%BB%93%E6%9E%9C%E4%B8%8D%E5%90%8C%E7%B1%BB%E7%B0%87%EF%BC%8C%E8%81%9A%E7%B1%BB%E7%BB%93%E6%9E%9C%E4%B8%8D%E5%90%8C%E7%B1%BB%E7%B0%87%20%20%0A#card=math&code=d%20%3D%20%7CDD%7C%2C%20DD%3D%5C%7B%28x_i%2Cx_j%29%7C%5Clambda_i%20%5Cneq%20%5Clambda_j%2C%5Clambda_i%5E%2A%20%5Cneq%20%5Clambda_j%5E%2A%2Ci%3Cj%20%5C%7D%20%EF%BC%8C%E5%8F%82%E8%80%83%E7%BB%93%E6%9E%9C%E4%B8%8D%E5%90%8C%E7%B1%BB%E7%B0%87%EF%BC%8C%E8%81%9A%E7%B1%BB%E7%BB%93%E6%9E%9C%E4%B8%8D%E5%90%8C%E7%B1%BB%E7%B0%87%20%20%0A&id=dnbZ1)

显然a和b代表着聚类结果好坏的正能量，b和c则表示参考结果和聚类结果相矛盾，基于这四个值可以导出以下常用的外部评价指标：

• Jaccard系统(Jaccard Coefficient，简称 JC)

• FM指数(Fowlkes and Mallows Index，简称FMI)

• Rand指数(Rand Index，简称RI)

%7D%7Bm(m-1)%7D%0A#card=math&code=RI%20%3D%20%5Cfrac%7B2%28a%2Bd%29%7D%7Bm%28m-1%29%7D%0A&id=m0c9E)

上述性能度量的结果值均在[0,1]区间，值越大越好。

内部指标

内部指标即不依赖任何外部模型，直接对聚类的结果进行评估，聚类的目的是想将那些相似的样本尽可能聚在一起，不相似的样本尽可能分开，直观来说：簇内高内聚紧紧抱团，簇间低耦合老死不相往来。定义：

%20%3D%20%20%5Cfrac%7B2%7D%7B%7CC%7C(%7CC%7C-1)%7D%5Csum%7Bi%5Cleq%20i%3Cj%20%5Cleq%7CC%7C%7Ddist(x_i%2Cx_j)%20%EF%BC%8C%E7%B0%87%E5%86%85%E6%B0%B4%E5%B9%B3%E8%B7%9D%E7%A6%BB%EF%BC%8C%E8%B6%8A%E5%B0%8F%E8%B6%8A%E5%A5%BD%20%0A#card=math&code=avg%28C%29%20%3D%20%20%5Cfrac%7B2%7D%7B%7CC%7C%28%7CC%7C-1%29%7D%5Csum%7Bi%5Cleq%20i%3Cj%20%5Cleq%7CC%7C%7Ddist%28x_i%2Cx_j%29%20%EF%BC%8C%E7%B0%87%E5%86%85%E6%B0%B4%E5%B9%B3%E8%B7%9D%E7%A6%BB%EF%BC%8C%E8%B6%8A%E5%B0%8F%E8%B6%8A%E5%A5%BD%20%0A&id=nTofv)

%20%3D%20max%7Bi%5Cleq%20i%20%3Cj%20%5Cleq%20%7CC%7C%7D%20dist(x_i%2Cx_j)%EF%BC%8C%E7%B0%87%E5%86%85%E6%9C%80%E5%A4%A7%E8%B7%9D%E7%A6%BB%EF%BC%8C%E8%B6%8A%E5%B0%8F%E8%B6%8A%E5%A5%BD%0A#card=math&code=diam%28C%29%20%3D%20max%7Bi%5Cleq%20i%20%3Cj%20%5Cleq%20%7CC%7C%7D%20dist%28x_i%2Cx_j%29%EF%BC%8C%E7%B0%87%E5%86%85%E6%9C%80%E5%A4%A7%E8%B7%9D%E7%A6%BB%EF%BC%8C%E8%B6%8A%E5%B0%8F%E8%B6%8A%E5%A5%BD%0A&id=Zayy6)

%20%3D%20min%7Bx_i%5Cin%20C_i%2Cx_j%20%5Cin%20C_j%7D%20dist(x_i%2Cx_j)%20%EF%BC%8C%E7%B0%87%E9%97%B4%E6%9C%80%E5%B0%8F%E8%B7%9D%E7%A6%BB%EF%BC%8C%E8%B6%8A%E5%A4%A7%E8%B6%8A%E5%A5%BD%0A#card=math&code=d%7Bmin%7D%28Ci%2CC_j%29%20%3D%20min%7Bx_i%5Cin%20C_i%2Cx_j%20%5Cin%20C_j%7D%20dist%28x_i%2Cx_j%29%20%EF%BC%8C%E7%B0%87%E9%97%B4%E6%9C%80%E5%B0%8F%E8%B7%9D%E7%A6%BB%EF%BC%8C%E8%B6%8A%E5%A4%A7%E8%B6%8A%E5%A5%BD%0A&id=ISbPy)

%3D%20dist(%5Cmui%2C%5Cmu_j)%EF%BC%8C%E7%B0%87%E4%B8%AD%E5%BF%83%E8%B7%9D%E7%A6%BB%EF%BC%8C%E8%B6%8A%E5%A4%A7%E8%B6%8A%E5%A5%BD%0A#card=math&code=d%7Bcen%7D%28C_i%2CC_j%29%3D%20dist%28%5Cmu_i%2C%5Cmu_j%29%EF%BC%8C%E7%B0%87%E4%B8%AD%E5%BF%83%E8%B7%9D%E7%A6%BB%EF%BC%8C%E8%B6%8A%E5%A4%A7%E8%B6%8A%E5%A5%BD%0A&id=mOugI)

基于上面的四个距离，可以导出下面这些常用的内部评价指标：

• DB 指数(Davies-Bouldin Index，简称DBI)

%2Bavg(Cj)%7D%7Bd%7Bcen%7D(%5Cmui%2C%5Cmu_j)%7D)%0A#card=math&code=DBI%20%3D%20%5Cfrac%7B1%7D%7Bk%7D%5Csum%7Bi%3D1%7D%5Ek%20%5Cunderset%7Bj%20%5Cneq%20i%7D%7Bmax%7D%28%5Cfrac%7Bavg%28Ci%29%2Bavg%28C_j%29%7D%7Bd%7Bcen%7D%28%5Cmu_i%2C%5Cmu_j%29%7D%29%0A&id=RmNWz)

• Dunn 指数(Dumm Index，简称DI)

%7D%7Bmax%7B1%5Cleq%20l%20%5Cleq%20k%7Ddiam(C_l)%7D)%20%5Cright%5C%7D%0A#card=math&code=DI%20%3D%20%5Cunderset%7B1%5Cleq%20i%20%5Cleq%20k%7D%7Bmin%7D%5Cleft%5C%7B%20%5Cunderset%7Bj%5Cneq%20i%7D%7Bmin%7D%28%5Cfrac%7Bd%7Bmin%7D%28Ci%2CC_j%29%7D%7Bmax%7B1%5Cleq%20l%20%5Cleq%20k%7Ddiam%28C_l%29%7D%29%20%5Cright%5C%7D%0A&id=h7UCK)

DBI的值越小越好，DI值越大越好。