个人笔记，非教程
对于每个样本 x i x_i xi，将其标记为距离类别中心最近的类别，即
l a b e l i = arg min ∣ ∣ x i − μ j ∣ ∣ label_i=\arg \min ||x_i-\mu_j|| labeli=argmin∣∣xi−μj∣∣
目标
min
S
S
E
=
∑
i
=
1
K
∑
x
j
∈
C
i
(
x
j
−
μ
i
)
2
=
∑
i
=
1
K
∑
x
j
∈
C
i
(
x
j
T
x
j
−
x
j
T
μ
i
−
μ
i
T
x
j
+
μ
i
T
μ
i
)
=
∑
i
=
1
K
(
∑
x
j
∈
C
i
x
j
T
x
j
−
∑
x
j
∈
C
i
x
j
T
μ
i
−
∑
x
j
∈
C
i
μ
i
T
x
j
+
∑
x
j
∈
C
i
μ
i
T
μ
i
)
=
∑
i
=
1
K
(
∑
x
j
∈
C
i
x
j
T
x
j
−
(
∑
x
j
∈
C
i
x
j
T
)
μ
i
−
μ
i
T
(
∑
x
j
∈
C
i
x
j
)
+
∣
C
i
∣
μ
i
T
μ
i
)
\begin{aligned} \min SSE&= \sum_{i=1}^{K}\sum_{x_j\in C_i} (x_j-\mu_i)^2 \\ &=\sum_{i=1}^{K}\sum_{x_j\in C_i} (x_j^Tx_j-x_j^T\mu_i-\mu_i^Tx_j+\mu_i^T\mu_i) \\ &=\sum_{i=1}^{K}(\sum_{x_j\in C_i} x_j^Tx_j-\sum_{x_j\in C_i}x_j^T\mu_i-\sum_{x_j\in C_i}\mu_i^Tx_j+\sum_{x_j\in C_i}\mu_i^T\mu_i) \\ &=\sum_{i=1}^{K}(\sum_{x_j\in C_i} x_j^Tx_j-(\sum_{x_j\in C_i}x_j^T)\mu_i-\mu_i^T(\sum_{x_j\in C_i}x_j)+|C_i|\mu_i^T\mu_i) \\ \end{aligned}
minSSE=i=1∑Kxj∈Ci∑(xj−μi)2=i=1∑Kxj∈Ci∑(xjTxj−xjTμi−μiTxj+μiTμi)=i=1∑K(xj∈Ci∑xjTxj−xj∈Ci∑xjTμi−xj∈Ci∑μiTxj+xj∈Ci∑μiTμi)=i=1∑K(xj∈Ci∑xjTxj−(xj∈Ci∑xjT)μi−μiT(xj∈Ci∑xj)+∣Ci∣μiTμi)
(SSE，误差平方和（Sum of the Squared Error,SSE）)
求导
∂ S S E ∂ μ i = − ( ∑ x j ∈ C i x j ) − ( ∑ x j ∈ C i x j ) + 2 ∣ C i ∣ μ i \frac{\partial SSE}{\partial \mu_i} =-(\sum_{x_j\in C_i}x_j)-(\sum_{x_j\in C_i}x_j)+2|C_i|\mu_i ∂μi∂SSE=−(xj∈Ci∑xj)−(xj∈Ci∑xj)+2∣Ci∣μi
令 ∂ S S E ∂ μ i = 0 \frac{\partial SSE}{\partial \mu_i} =0 ∂μi∂SSE=0
μ i = 1 ∣ C i ∣ ∑ x j ∈ C i x j \mu_i=\frac{1}{|C_i|} \sum_{x_j\in C_i}x_j μi=∣Ci∣1xj∈Ci∑xj
求解，反复迭代即可
l
a
b
e
l
i
=
arg
min
∣
∣
x
i
−
μ
j
∣
∣
label_i=\arg \min ||x_i-\mu_j||
labeli=argmin∣∣xi−μj∣∣
μ
i
=
1
∣
C
i
∣
∑
x
j
∈
C
i
x
j
\mu_i=\frac{1}{|C_i|} \sum_{x_j\in C_i}x_j
μi=∣Ci∣1xj∈Ci∑xj
254
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
