Matlab利用子函数EuDist2函数，构建近邻图，EuDist2函数为构建欧几里得距离图

在学习关于如何构建下图公式所示近邻图时，有一个非常简答的调用函数：EuDist2

下面有两种代码实现效果一样。

第一种：简单直接

clear all
clc
%% method 1 
fea = rand(100,200);
num_sample = size(fea,1);
class_num = size(fea,1)/10;
tic
Ww = zeros(num_sample,num_sample);
Wb = ones(num_sample,num_sample);
G = zeros(num_sample,num_sample);
for i = 1:class_num
    gnd((i-1)*10+1:i*10) = i;
    Ww((i-1)*10+1:i*10,(i-1)*10+1:i*10) = 1;
    Wb((i-1)*10+1:i*10,(i-1)*10+1:i*10) = 0;
end
k = 5;
for i = 1:num_sample
    for j = 1:num_sample
        F(i,j) = norm(fea(i,:)-fea(j,:));
    end
    [aa,bb] = sort(F(i,:));
    for kk = 1:k
        if gnd(bb(kk)) == gnd(i)
            G(i,bb(kk)) = 1;
        end
    end
end
Ww = Ww.*G;
Wb = Wb.*G;
t1 = toc

第二种，快速，还用到了sparse稀疏矩阵，哎呀，理解比较麻烦，不过速度确实比第一种快

clear all
clc
%% method 1 
fea = rand(100,200);
num_sample = size(fea,1);
class_num = size(fea,1)/10;
tic
uniqueLabels = unique(gnd);
Ww1 = zeros(num_sample,num_sample);
Wb1 = ones(num_sample,num_sample);
for idx = 1 : class_num
    classIdx = find(gnd == uniqueLabels(idx));
    Ww1(classIdx,classIdx) = 1;
    Wb1(classIdx,classIdx) = 0;
end

distMat = EuDist2(fea,[],0);
[dump sortedIDs] = sort(distMat,2);             % 逐行升序排列
clear dump
kIDs = sortedIDs(:,1 : k);

G1 = sparse(repmat([1 : num_sample]',[k,1]),kIDs(:),ones(prod(size(kIDs)),1),num_sample,num_sample);
G1 = full(max(G1,G1'));
Ww1 = full(Ww .* G1);
Wb1 = full(Wb .* G1);
t2 = toc

function D = EuDist2(fea_a,fea_b,bSqrt)
%EUDIST2 Efficiently Compute the Euclidean Distance Matrix by Exploring the
%Matlab matrix operations.
%
%   D = EuDist(fea_a,fea_b)
%   fea_a:    nSample_a * nFeature
%   fea_b:    nSample_b * nFeature
%   D:      nSample_a * nSample_a
%       or  nSample_a * nSample_b
%
%    Examples:
%
%       a = rand(500,10);
%       b = rand(1000,10);
%
%       A = EuDist2(a); % A: 500*500
%       D = EuDist2(a,b); % D: 500*1000
%
%   version 2.1 --November/2011
%   version 2.0 --May/2009
%   version 1.0 --November/2005
%
%   Written by Deng Cai (dengcai AT gmail.com)


if ~exist('bSqrt','var')
    bSqrt = 1;
end

if (~exist('fea_b','var')) || isempty(fea_b)
    aa = sum(fea_a.*fea_a,2);
    ab = fea_a*fea_a';
    
    if issparse(aa)
        aa = full(aa);                  % transmit the sparse matrix into the full matrix
    end
    
    D = bsxfun(@plus,aa,aa') - 2*ab;
    D(D<0) = 0;
    if bSqrt
        D = sqrt(D);
    end
    D = max(D,D');
else
    aa = sum(fea_a.*fea_a,2);
    bb = sum(fea_b.*fea_b,2);
    ab = fea_a*fea_b';

    if issparse(aa)
        aa = full(aa);
        bb = full(bb);
    end

    D = bsxfun(@plus,aa,bb') - 2*ab;
    D(D<0) = 0;
    if bSqrt
        D = sqrt(D);
    end
end

运行时间比较：

t1 =

    0.0221

t2 =

    0.0189

应该矩阵越大的话，肯定加速更明显吧。