classify and crossval function which bad with positive definite

1 I have a set of data (20 x 1200)(sample x variance) (1)the first colum of data is the group data [1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 3 3 3 3 3 ] thus the group 1 has 7 samples,group 2 has 8 samples, group 3 has 5 samples,

(2)the rest colum (2:end) is the specrta data.

2 (1)when I use classify function with tpye of linear,quadratic, mahalanobis in matlab 2012

It runs error as below:

its faults classify (line 277)
The covariance matrix of each group in TRAINING must be positive definite.

(2)when I use classify function with tpye of diaglinear (or diagquadratic )its OK

and then I make crossvalidation with code of:

    cp = cvpartition(group,'leaveout');
    ldaClassFun= @(xtrain,ytrain,xtest)(classify(xtest,xtrain,ytrain));
    ldaCVErr  = crossval('mcr',training,group,'predfun',ldaClassFun,'partition',cp)

It runs error as below:

its faults crossval>evalFun (line 480)
The function '@(xtrain,ytrain,xtest)(classify(xtest,xtrain,ytrain))' generated the following error:
The pooled covariance matrix of TRAINING must be positive definite.

its faults crossval>getLossVal (line 517)
funResult = evalFun(funorStr,arg(1:end-1));

its faults crossval (line 416)
                [funResult,outarg] = getLossVal(i, nData, cvp, data, predfun);

So, did anyone know why and how to cope with them!!! Thanks very much!!!

Answer 1

[Roger Wohlwend]

Matlab demands a positive definite covariance matrix and in your case that requirement is obviously not met. What can you do?

• If your data is not stationary, then using stationary data (i.e. using differences instead of levels) could solve the problem.
• Check if some columns in your data are linearly dependent on each other.
• If the columns have different scales (for example one column consists of very small numbers, another columns has very large numbers), scale the columns so that the numbers are approximately of the same magnitude