Undefined function 'mtimes' for input arguments of type 'nominal'.

Dear all,

I want to use cross-validation function crossval to find some parameters of my SVM function. I follow the help document to have this .m file:

load UCI_abalone; %2088
data = [train_data;test_data];
label = [train_labels;test_labels];
data_size = size(data,1);

c = cvpartition(data_size,'kfold',10);
minfn = @(z)crossval('mcr',data,label,'Predfun', ...
    @(xtrain,ytrain,xtest)crossfun(xtrain,ytrain,...
    xtest,exp(z)),'partition',c);

opts = optimset('TolX',5e-4,'TolFun',5e-4);
[searchmin fval] = fminsearch(minfn,randn(1,1),opts)

the corresponding function is:

function [ ytest ] = crossfun(x,y,xtest,lambda)
p = size(x,2);
N = size(x,1);
w = zeros(p,1);
for i = 1:N
    yita = 1 / sqrt(i);
    if 1 - y(i) * x(i,:) * w  > 0
        w = w + y(i) * yita * x(i,:)';
    end
    if w' * w > 1 / lambda
        w = w / sqrt(w' * w) / lambda;
    end
end
ytest = xtest * w;
end

And when I run it, the error shows like this:

>> cv_test

Error using crossval>evalFun (line 465)
The function '@(xtrain,ytrain,xtest)crossfun(xtrain,ytrain,xtest,exp(z))' generated the following error:
Undefined function 'mtimes' for input arguments of type 'nominal'.

Error in crossval>getLossVal (line 502)
funResult = evalFun(funorStr,arg(1:end-1));

Error in crossval (line 401)
                [funResult,outarg] = getLossVal(i, nData, cvp, data, predfun);

Error in
@(z)crossval('mcr',data,label,'Predfun',@(xtrain,ytrain,xtest)crossfun(xtrain,ytrain,xtest,exp(z)),'partition',c)

Error in fminsearch (line 191)
fv(:,1) = funfcn(x,varargin{:});

Error in cv_test (line 12)
[searchmin fval] = fminsearch(minfn,randn(1,1),opts)

Is there anyone can help me??? Thanks!!

Answer 1

[Tom Lane]

Try putting a break in the crossfun function and using "whos" to see what the y value is. If it's a nominal variable, then it looks like the problem is that you are trying to multiply it by a row of the x matrix.

I'm not exactly sure what you are trying to do. Could it be that you want the "label" variable to take on values +1 for one class and -1 for the other?

[Yang]

Yeah you're right. This is a classification problem. "label" variable takes +1 and -1 corresponding to different feature data. I want to use cross-validation to find out an optimal value of lambda which is a regularization parameter in SVM approach. I will try your advise and give you response later. Thanks!

[Yang]

I tried your suggest. The parameter "label" that I passed is double type. But in some functions it was convert to the type "nominal" but still take the value of +1 and -1. I don't know why and how to solve that.

[Jan]

@Yang: Because we neither know where and also not why the value is changed, we cannot know how to solve this either.

[Tom Lane]

I also don't know where it changed. I can imagine some Statistics Toolbox code may operate on nominal values internally, but I would not expect that to affect you. A possibility is to write your crossfun function to deal with nominals. For instance, double(y(i)) will give 1 for y's first category, 2 for y's second category, and so on.