论文欣赏—A Linear Assignment Clustering Algorithm Based on the Least Similar Cluster Representatives

这篇博文列出了主要的公式，和讲解视频一起看效果更佳！
讲解视频

Similarity and Dissimilarity Coefficients——相似度不相似度的定义

首先我们来看一下两个系数的定义方法，Similarity and dissimilarity coefficients代表着两个数据之间的相似和不相似的程度
一个典型的dissimilarity coefficient是the Minkowski metric
$$d_{i,j}=[\sum _{k=1}^m(a_{ki}-a_{kj}{)}^q{]}^{\frac{1}{q}}$$
其中$q>0$
定义similarity coefficient
$$s_{ij}=\frac{{\sum }_{k=1}^m{a}_{ki}{a}_{kj}}{{\sum }_{k=1}^m(a_{ki}+a_{kj}-a_{ki}{a}_{kj})}$$

Cluster Representatives——选代表

方法一：
{ r 1 , r 2 } = a r g min ⁡ ( i , j ) s i j \{r_1,r_2\}=arg\min_{(i,j)}s_{ij} {r1​,r2​}=arg(i,j)min​sij​
r k = a r g min ⁡ i ∈ { 1 , 2 , . . . , k − 1 } ∑ j = 1 k − 1 s i r j k = 3 , 4 , . . . p r_k=arg\min_{i\in\{1,2,...,k-1\}}\sum_{j=1}^{k-1}s_ir_j\\ k=3,4,...p rk​=argi∈{1,2,...,k−1}min​j=1∑k−1​si​rj​k=3,4,...p
其中$r_k$表示的是第k个cluter的index
方法二：
{ r 1 , r 2 . . . r p } = a r g min ⁡ r ∈ { 1 , 2 , . . . p } { ∑ i = 1 n ∑ r < i s r i r j ∣ r i , r j ∈ { 1 , 2 , . . . , n } } \{r_1,r_2...r_p\}=arg\min_{r\in\{1,2,...p\}}\bigg\{\sum_{i=1}^n\sum_{r<i}s_{r_ir_j}|r_i,r_j\in\{1,2,...,n\}\bigg\} {r1​,r2​...rp​}=argr∈{1,2,...p}min​{i=1∑n​r<i∑​sri​rj​​∣ri​,rj​∈{1,2,...,n}}

Linear Assignment Model

$$Maximize\sum _{i=1}^n\sum _{k=1}^p{s}_{i{r}_k}{x}_{ik}orminimize\sum _{i=1}^n\sum _{k=1}^p{d}_{i{r}_k}{x}_{ik}$$
$$subjectto\sum _{i=1}^n{x}_{ik}=1k=1,2,...p$$
$$\sum _{k=1}^p{x}_{ik}\le ui=1,2,...n$$
$$x_{ik}\ge 0i=1,2,...,n;k=1,2,...,p$$
其中$x_{ik}$是二值决策变量

Assignment Clustering Algorithm

Step 0: Set I = { i ∣ 1 , 2 , . . . , n } I=\{i|1,2,...,n\} I={i∣1,2,...,n}and K = { k ∣ 1 , 2 , . . . p } . K=\{k|1,2,...p\}. K={k∣1,2,...p}.
Step 1: Load the number of clusters n and the upper bound of data per cluster u:
Step 2: Load or compute the similarity coefficients between every pair of data.
Step 3: Determine cluster representatives using (15) and (16) or (17), then remove $r_k(k|1,2,...p)$from I.
Step 4: Determine $(v,w)=arg{\max }_{i\in I,k\in K}{s}_{i{r}_k}$.
Step 5: If the number of data in cluster w is u; then remove w from K and go to Step 4; otherwise, assign datum v to cluster w and delete v from I:
Step 6: If $I\ne \varnothing$; go to Step 4.
Step 7: Evaluate the clustering result using one or more performance criteria.