欢迎您访问程序员文章站本站旨在为大家提供分享程序员计算机编程知识!
您现在的位置是: 首页

Matlab实现Kmeans聚类,并利用匈牙利算法Kuhn-Munkres实现对聚类标签和真实标签的映射,对结果进行聚类精度Accuracy评价和标准互信息Nmi评价

程序员文章站 2022-07-14 20:18:31
...

思路

  1. 利用Kmeans算法对数据进行聚类,生成聚类结果
  2. 将聚类结果的标签和真实标签进行映射,生成映射后标签
  3. 利用映射后标签和真实标签进行计算Accuracy值
  4. 利用聚类结果标签和真实标签计算Nmi值

输入数据为经典MNIST数据集

MNIST数据集是机器学习领域中非常经典的一个数据集,由60000个训练样本和10000个测试样本组成,每个样本都是一张28 * 28像素的灰度手写数字图片。
我们选取MNIST数据集中的10类共3000条数据为例。

  • 数据下载网址:http://yann.lecun.com/exdb/mnist/

利用Matlab中Kmeans算法对数据进行聚类

clc,clear;
load E:\2019\机器学习\实验一\data\MNIST;
opts = statset('Display','final');
K=10;           %将X划分为K类
repN=50;        %迭代次数
%K-mean聚类
[Idx,Ctrs,SumD,D] = kmeans(X,K,'Replicates',repN,'Options',opts);

聚类精度Acc评价

聚类精度(Acc):给定一个聚类结果标签 和其对应的指示标签 ,Acc计算公式如下:
Matlab实现Kmeans聚类,并利用匈牙利算法Kuhn-Munkres实现对聚类标签和真实标签的映射,对结果进行聚类精度Accuracy评价和标准互信息Nmi评价
其中:
Matlab实现Kmeans聚类,并利用匈牙利算法Kuhn-Munkres实现对聚类标签和真实标签的映射,对结果进行聚类精度Accuracy评价和标准互信息Nmi评价Matlab实现Kmeans聚类,并利用匈牙利算法Kuhn-Munkres实现对聚类标签和真实标签的映射,对结果进行聚类精度Accuracy评价和标准互信息Nmi评价
map(oi)是一个映射函数,它以真实标签 gi作为参考标签,然后按照相同的排列方式oi中的标签顺序进行重排。因此, map(oi)是用来解决标签不一致问题。通常可采用经典的Kuhn-Munkres算法实现 的重排。

标准互信息NMI评价

标准互信息(NMI):互信息(MI)是一种堆成的度量方式,他可以衡量两种分布之间相互依赖程度,判断两种分布的一致性。设 cp表示真实标签 c中的第gi 类,c’q表示oi中的第 q类,则对应的MI可定义为:
Matlab实现Kmeans聚类,并利用匈牙利算法Kuhn-Munkres实现对聚类标签和真实标签的映射,对结果进行聚类精度Accuracy评价和标准互信息Nmi评价
其中k和k’分别表示真实标签和聚类标签对应的类别数。np表示类别cp包含的样本数,n’q表示类别c’q中包含的样本数,npq表示同时出现在类别cp和c’q中的样本数,那么标准互信息可以定义为:
Matlab实现Kmeans聚类,并利用匈牙利算法Kuhn-Munkres实现对聚类标签和真实标签的映射,对结果进行聚类精度Accuracy评价和标准互信息Nmi评价
其中H(g)是熵函数。根据上述两式,有:
Matlab实现Kmeans聚类,并利用匈牙利算法Kuhn-Munkres实现对聚类标签和真实标签的映射,对结果进行聚类精度Accuracy评价和标准互信息Nmi评价

利用Kuhn-Munkres算法实现munkres方法

匈牙利算法很经典,一般用来解决最优分配问题,这里从Matlab论坛上找到一个比较好的实现,网上有很多帖子对匈牙利算法的原理讲的都很不错,这里推荐一个详细官方英文文档讲解:匈牙利算法官网英文文档

function [assignment] = munkres(costMat)
% MUNKRES   Munkres Assign Algorithm
%
% [ASSIGN,COST] = munkres(COSTMAT) returns the optimal assignment in ASSIGN
% with the minimum COST based on the assignment problem represented by the
% COSTMAT, where the (i,j)th element represents the cost to assign the jth
% job to the ith worker.
%
 
% This is vectorized implementation of the algorithm. It is the fastest
% among all Matlab implementations of the algorithm.
 
% Examples
% Example 1: a 5 x 5 example
%{
[assignment,cost] = munkres(magic(5));
[assignedrows,dum]=find(assignment);
disp(assignedrows'); % 3 2 1 5 4
disp(cost); %15
%}
% Example 2: 400 x 400 random data
%{
n=5;
A=rand(n);
tic
[a,b]=munkres(A);
toc                
%}
 
% Reference:
% "Munkres' Assignment Algorithm, Modified for Rectangular Matrices",
% http://csclab.murraystate.edu/bob.pilgrim/445/munkres.html
 
% version 1.0 by Yi Cao at Cranfield University on 17th June 2008
 
assignment = false(size(costMat));
 
costMat(costMat~=costMat)=Inf;
validMat = costMat<Inf;
validCol = any(validMat);
validRow = any(validMat,2);
 
nRows = sum(validRow);
nCols = sum(validCol);
n = max(nRows,nCols);
if ~n
    return
end
     
dMat = zeros(n);
dMat(1:nRows,1:nCols) = costMat(validRow,validCol);
 
%*************************************************
% Munkres' Assignment Algorithm starts here
%*************************************************
 
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%   STEP 1: Subtract the row minimum from each row.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
 dMat = bsxfun(@minus, dMat, min(dMat,[],2));
 
%************************************************************************** 
%   STEP 2: Find a zero of dMat. If there are no starred zeros in its
%           column or row start the zero. Repeat for each zero
%**************************************************************************
zP = ~dMat;
starZ = false(n);
while any(zP(:))
    [r,c]=find(zP,1);
    starZ(r,c)=true;
    zP(r,:)=false;
    zP(:,c)=false;
end
 
while 1
%**************************************************************************
%   STEP 3: Cover each column with a starred zero. If all the columns are
%           covered then the matching is maximum
%**************************************************************************
    primeZ = false(n);
    coverColumn = any(starZ);
    if ~any(~coverColumn)
        break
    end
    coverRow = false(n,1);
    while 1
        %**************************************************************************
        %   STEP 4: Find a noncovered zero and prime it.  If there is no starred
        %           zero in the row containing this primed zero, Go to Step 5. 
        %           Otherwise, cover this row and uncover the column containing
        %           the starred zero. Continue in this manner until there are no
        %           uncovered zeros left. Save the smallest uncovered value and
        %           Go to Step 6.
        %**************************************************************************
        zP(:) = false;
        zP(~coverRow,~coverColumn) = ~dMat(~coverRow,~coverColumn);
        Step = 6;
        while any(any(zP(~coverRow,~coverColumn)))
            [uZr,uZc] = find(zP,1);
            primeZ(uZr,uZc) = true;
            stz = starZ(uZr,:);
            if ~any(stz)
                Step = 5;
                break;
            end
            coverRow(uZr) = true;
            coverColumn(stz) = false;
            zP(uZr,:) = false;
            zP(~coverRow,stz) = ~dMat(~coverRow,stz);
        end
        if Step == 6
            % *************************************************************************
            % STEP 6: Add the minimum uncovered value to every element of each covered
            %         row, and subtract it from every element of each uncovered column.
            %         Return to Step 4 without altering any stars, primes, or covered lines.
            %**************************************************************************
            M=dMat(~coverRow,~coverColumn);
            minval=min(min(M));
            if minval==inf
                return
            end
            dMat(coverRow,coverColumn)=dMat(coverRow,coverColumn)+minval;
            dMat(~coverRow,~coverColumn)=M-minval;
        else
            break
        end
    end
    %**************************************************************************
    % STEP 5:
    %  Construct a series of alternating primed and starred zeros as
    %  follows:
    %  Let Z0 represent the uncovered primed zero found in Step 4.
    %  Let Z1 denote the starred zero in the column of Z0 (if any).
    %  Let Z2 denote the primed zero in the row of Z1 (there will always
    %  be one).  Continue until the series terminates at a primed zero
    %  that has no starred zero in its column.  Unstar each starred
    %  zero of the series, star each primed zero of the series, erase
    %  all primes and uncover every line in the matrix.  Return to Step 3.
    %**************************************************************************
    rowZ1 = starZ(:,uZc);
    starZ(uZr,uZc)=true;
    while any(rowZ1)
        starZ(rowZ1,uZc)=false;
        uZc = primeZ(rowZ1,:);
        uZr = rowZ1;
        rowZ1 = starZ(:,uZc);
        starZ(uZr,uZc)=true;
    end
end
%生成标签矩阵
assignment(validRow,validCol) = starZ(1:nRows,1:nCols);

%解决标签映射问题不需要计算权重cost,故将其注释
%cost = 0;
%cost = sum(costMat(assignment));

对Kmeans聚类结果进行映射

function [NewLabel] = BestMapping(La1,La2)

%真实标签:La1 聚类结果标签:La2 映射后的标签:NewLabel

Label1=unique(La1');
L1=length(Label1);
Label2=unique(La2');
L2=length(Label2);

%构建计算两种分类标签重复度的矩阵G
G = zeros(max(L1,L2),max(L1,L2));
for i=1:L1
    index1= La1==Label1(1,i);
    for j=1:L2
        index2= La2==Label2(1,j);
        G(i,j)=sum(index1.*index2);
    end
end

%利用匈牙利算法计算出映射重排后的矩阵
[index]=munkres(-G);
%将映射重排结果转换为一个存储有映射重排后标签顺序的行向量
[temp]=MarkReplace(index);
%生成映射重排后的标签NewLabel
NewLabel=zeros(size(La2));
for i=1:L2
    NewLabel(La2==Label2(i))=temp(i);
end

end

聚类结果和真实标签进行映射的核心思路

设真实标签有m类,聚类结果标签有n类,L=max(m,n)。生成一个大小为L*L且元素均为0的矩阵。本文中的聚类结果类别数目和真实标签类别数目一样,即m=n。计算真实标签和聚类标签结果的重复度,并将结果存储在矩阵G中,这个计算过程体现在以下代码里:

for i=1:L1
    index1= La1==Label1(1,i);
    for j=1:L2
        index2= La2==Label2(1,j);
        G(i,j)=sum(index1.*index2);
    end
end

其中循环结构中分别将真实标签、聚类结果标签中相同类别的点的分布用01列矩阵表示了出来。然后对两个01列矩阵进行点乘求和运算,下面画了一个示意图:
Matlab实现Kmeans聚类,并利用匈牙利算法Kuhn-Munkres实现对聚类标签和真实标签的映射,对结果进行聚类精度Accuracy评价和标准互信息Nmi评价
可以看到相同分布的地方经过点乘后会得到1,不同分布的地方点乘后得到0,这样点乘后的矩阵里面的1表示聚类标签和真实标签相同重复的地方,对这一列求和得到的数字就是聚类结果和真实结果的重复度,将这个结果记录在矩阵G里,同理依次求出聚类结果标签m类和真实标签n类的m*n个重复度,分别记录在G中。
按照思路我们直接找出G中每一行中重复度最大的值,确定它的位置就可以了,但是这样的话会出现多个不同行的重复度最大的值在同一列的情况(即真实标签和聚类结果标签的映射不是1对1),这显然是不合理的。
我们发现这个问题其实就是一个最佳分配问题,所以可以利用匈牙利算法解决。这样就可以成功找到真实标签和聚类结果标签1对1的映射。

将存储点分布的空间矩阵转换为一个行向量的辅助方法

%将存储标签顺序的空间矩阵转换为一个行向量
function [assignment] = MarkReplace(MarkMat)

[rows,cols]=size(MarkMat);

assignment=zeros(1,cols);

for i=1:rows
    for j=1:cols
        if MarkMat(i,j)==1
            assignment(1,j)=i;
        end
    end
end

end

Acc计算实现

function acc = Acc(Label1,Label2)
%Label1:真实标签 Label2:映射后的标签

T= Label1==Label2;
acc=sum(T)/length(Label2);

end

Nmi计算实现

function nmi = Nmi(A,B)
%A:真实标签 B:聚类标签

%NMI Normalized mutual information
% http://en.wikipedia.org/wiki/Mutual_information
% http://nlp.stanford.edu/IR-book/html/htmledition/evaluation-of-clustering-1.html

if length( A ) ~= length( B)
    error('length( A ) must == length( B)');
end
total = length(A);
A_ids = unique(A);
A_class = length(A_ids);
B_ids = unique(B);
B_class = length(B_ids);
% Mutual information
idAOccur = double (repmat( A, A_class, 1) == repmat( A_ids', 1, total ));
idBOccur = double (repmat( B, B_class, 1) == repmat( B_ids', 1, total ));
idABOccur = idAOccur * idBOccur';
Px = sum(idAOccur') / total;
Py = sum(idBOccur') / total;
Pxy = idABOccur / total;
MImatrix = Pxy .* log2(Pxy ./(Px' * Py)+eps);
MI = sum(MImatrix(:));
% Entropies
Hx = -sum(Px .* log2(Px + eps),2);
Hy = -sum(Py .* log2(Py + eps),2);
%Normalized Mutual information
nmi = 2 * MI / (Hx+Hy);

% Nmi = MI / sqrt(Hx*Hy); another version of NMI

end


Kmeans聚类并评价

clc,clear;
load E:\2019\机器学习\实验一\data\MNIST;

K=10;           %将X划分为K类
repN=50;        %迭代次数
opts = statset('Display','final');

%K-mean聚类
[Idx,Ctrs,SumD,D] = kmeans(X,K,'Replicates',repN,'Options',opts);

%打印结果
fprintf('划分成%d类的结果如下:\n',K)
for i=1:K
    tm=find(Idx==i); %求第i类的对象
    tm=reshape(tm,1,length(tm)); %变成行向量
    fprintf('第%d类共%d个分别是%s\n',i,length(tm),int2str(tm)); %显示分类结果
end

%进行映射操作
[NewLabel]=BestMapping(Y,Idx);
 
%Y:真实标签 Idx:聚类标签 NewLabel:映射重排后的标签

%ACC
acc=Acc(Y,NewLabel);
fprintf('聚类的精度Acc为:%f\n',acc); %显示分类结果

%NMI
nmi=Nmi(Y',Idx');
fprintf('聚类的标准互信息Nmi为:%f\n',nmi); %显示分类结果

运行结果

Matlab实现Kmeans聚类,并利用匈牙利算法Kuhn-Munkres实现对聚类标签和真实标签的映射,对结果进行聚类精度Accuracy评价和标准互信息Nmi评价