Q13-15 rbf vsm eout=0?

[zuodongz]

Q13-15 EOUT from rbf support vector machine is always 0.and this machine can never be beaten by k mean rbf .... what is the mistake i probably made?

[yaser]

Originally Posted by zuodongz

Q13-15 EOUT from rbf support vector machine is always 0.and this machine can never be beaten by k mean rbf .... what is the mistake i probably made?

Why is always zero?

BTW, if your reply would reveal answers (or excluded answers) of a question, please start an *ANSWER* thread for that (see above announcement).

[zuodongz]

Originally Posted by yaser

Why is always zero?

BTW, if your reply would reveal answers (or excluded answers) of a question, please start an *ANSWER* thread for that (see above announcement).

Hi professor
Sorry for the troubles i made... and thanks for your quick reply
The Eout I achieved from RBF kernel vector support machine is always 0. Is correct ?
you gave excellent lectures! What book would you recommend for further studying on machine learning , or machine learning in financial field?

[yaser]

Originally Posted by zuodongz

The Eout I achieved from RBF kernel vector support machine is always 0. Is correct ?

Are you estimating based on a fresh sample of points that were not involved in training? If so, is the sample you are using big enough to estimate a small, but non-zero value?

you gave excellent lectures! What book would you recommend for further studying on machine learning , or machine learning in financial field?

Thank you. After you are done with the course and the e-chapters on this forum, you are ready to start trying what you have learned on real data, and ready to read current papers on Machine Learning to update your techniques.

[zuodongz]

Originally Posted by yaser

Are you estimating based on a fresh sample of points that were not involved in training? If so, is the sample you are using big enough to estimate a small, but non-zero value?


Thank you. After you are done with the course and the e-chapters on this forum, you are ready to start trying what you have learned on real data, and ready to read current papers on Machine Learning to update your techniques.

Hi professor
The training data and testing data are totally separate. I used 100 training data to train this rbf vector support machine ,and randomly generate 100 new points for testing. (no matter how many times i randomly generate 100 points , Eout is always 0)

[yaser]

Originally Posted by zuodongz

Hi professor
The training data and testing data are totally separate. I used 100 training data to train this rbf vector support machine ,and randomly generate 100 new points for testing. (no matter how many times i randomly generate 100 points , Eout is always 0)

Try 10,000 points for testing. Since you say that the error is zero no matter how many times you generated new 100 points, you may also want to check the seeding of the random number generator to make sure that each 100 points are really different from the previous set.

A zero out-of-sample error means that the target boundary was perfectly replicated, an illogical event when we train with only 100 points.

[zuodongz]

Originally Posted by yaser

Try 10,000 points for testing. Since you say that the error is zero no matter how many times you generated new 100 points, you may also want to check the seeding of the random number generator to make sure that each 100 points are really different from the previous set.

A zero out-of-sample error means that the target boundary was perfectly replicated, an illogical event when we train with only 100 points.

Thanks!! professor .

[zuodongz]

Originally Posted by yaser

Try 10,000 points for testing. Since you say that the error is zero no matter how many times you generated new 100 points, you may also want to check the seeding of the random number generator to make sure that each 100 points are really different from the previous set.

A zero out-of-sample error means that the target boundary was perfectly replicated, an illogical event when we train with only 100 points.

I think i still have problems with kernel vector machine, as I calculated alpha from quadratic programming , which is (N*1) matrix, Then I calculated b by using
MATLAB CODE %%%%
[maxval, maxind] = max(alpha);
b=1/y(maxind)- kernel_x(maxind,*(alpha.*y);
%%%%
and i was unable to get w , then Ein was measured by
%%%
pred=sign(kernel_x*alpha.*y+b);
Ein_svm=numel(find(pred~=y));

For test , i used

pred=sign(test_kernel_x*alpha.*y+b);
Eout_svm=numel(find(pred~=y_test)).

as if the number of test data points are not consistent with number of training data points. test_kernel_x*alpha.*y is unable to be calculated , as dimensions do not match.......

I think there is some confusion about my understanding to kernel vector machine.
isn't w unachievable? Thanks professor.

[zuodongz]

correction *b=1/y(maxind)- kernel_x(maxind,)*(alpha.*y);