基于matlab的svdd多分类程序,SVDD算法程序解释的问题？
function W = svdd(a,fracrej,sigma)
% First set up the parameters
if nargin < 3
sigma = [];
end
if nargin < 2 || isempty(fracrej), fracrej = 0.05; end
if nargin < 1 || isempty(a) % empty svdd
W = mapping(mfilename,{fracrej,sigma});
W = setname(W,'Support vector data description');
return
end
if isempty(sigma),
sigma = 5;
end
if ~ismapping(fracrej) % training
if isempty(sigma)
error('This versions needs a sigma.');
end
% introduce outlier label for outlier class if it is available.
if isocset(a) %检查数据是否是一类数据
signlab = getoclab(a); %从数据a中得到数字标签
if all(signlab<0), error('SVDD needs target objects!'); end %如果所有的标签小于零，那么svdd出错！！需要目标样本
else
%error('SVDD needs a one-class dataset.');
% Noo, be nice, everything is target:
signlab = ones(size(a,1),1);
%a = target_class(+a);
end
% check the rejection rates % 检查拒绝率
if (length(fracrej)>2) % if no bound on the outlier error is given, we
if length(fracrej)~=length(signlab)
error('The length of C is not fitting.');
end
C = fracrej;
else
if (length(fracrej)<2) % if no bound on the outlier error is given, we
% do not care
fracrej(2) = 1;
end
if (fracrej(1)>1)
warning('dd_tools:AllReject',...
'Fracrej > 1? I cannot reject more than all my target data!');
end
fracrej
% Setup the appropriate C's %设置c值
if (fracrej(1)<0) %DXD: tricky trick....
C(1) = -fracrej(1);
C(2) = -fracrej(2);
else
nrtar = length(find(signlab==1));
nrout = length(find(signlab==-1));
% we could get divide by zero, but that is ok.
%warning off MATLAB:divideByZero;
C(1) = 1/(nrtar*fracrej(1));
C(2) = 1/(nrout*fracrej(2));%nrout 异常样本数
%warning on MATLAB:divideByZero;
end
end
% Find the alpha's
% Standard optimization procedure:
[alf,R2,Dx,J] = svdd_optrbf(sigma,+a,signlab,C); %二次优化 J为支持向量和错误的点
SVx = +a(J,:);
alf = alf(J);
% Compute the offset (not important, but now gives the possibility to 计算偏移量(不重要,但是现在提供了可能性
% interpret the output as the distance to the center of the sphere) 解释输出与球体的中心的距离)
offs = 1 + sum(sum((alf*alf').*exp(-sqeucldistm(SVx,SVx)/(sigma*sigma)),2)) %sqeucldistm专门计算欧式距离
% store the results %保存结果
W.c =C;
W.s = sigma;
W.a = alf;
W.R2= R2;
W.threshold = offs+R2;
W.sv = SVx;
W.offs = offs;
W = mapping(mfilename,'trained',W,char('target','outlier'),size(a,2),2);
W = setname(W,'Support vector data description');
else %testing
W = getdata(fracrej);
m = size(a,1)
% check if alpha's are OK 检查α是否合适
if isempty(W.a)
warning('dd_tools:OptimFailed','The SVDD is empty or not well defined');
out = zeros(m,1);
else
% and here we go:
K = exp(-sqeucldistm(+a,W.sv)/(W.s*W.s)); %欧氏距离的平方
out = W.offs - 2*sum( repmat(W.a',m,1).* K, 2);
end
newout = [out repmat(W.threshold,m,1)];
% Store the distance as output:
W = setdat(a,-newout,fracrej); %重置数据和功能标签的数据集分类输出
W = setfeatdom(W,{[-inf 0; -inf 0] [-inf 0; -inf 0]});
end
return
搞不清楚training和testing是做什么的？newout求出的是什么？