Difference result of SMO and LibSVM for same data

[Saeedeh]

Hi

I used smo and libsvm in weka to classify data. but the results are different.

the data that used in both of the classifier is same, but smo have better result. smo  accuracy is 7 percent higher than libsvm. 

I though that libsvm use smo and the result should be same.

why the results are different?

[Bhoomin Pandya]

Hello

I suggest you must review some of the details about the implements posted on the

http://list.waikato.ac.nz/  & http://weka.wikispaces.com/

they should provide you a lot of insight into your specific problem you are facing.

Also please see

eg. http://comments.gmane.org/gmane.comp.ai.weka/34107

Thanks,
Bhoomin Pandya

[Bhoomin Pandya]

Hello Saeedah:

I am curious about installation of LibSVM, LibLINEAR and WLSVM. Could you give me some hints. I downloaded the WLSVM from http://www.cs.iastate.edu/~yasser/wlsvm/. I am just running WEKA 3.6.11 stable (book) version.
I am getting error initiating these Classfiers due to even after I set classpath for LibSVM and WLSVM jars with error. I believe there is a separate Package Manager UI also which I have to explore.

Regarding your problem you would also want to go through some of Dr. Ian Witten's Class videos. (I am not too sure if Dr. Ian Witten had addressed LibSVM classifiers in one of his class videos from the most recent course More Data Mining with WEKA. If I trace it I shall let you know.) 

On Sunday, June 8, 2014 4:13:23 PM UTC+5:30, Saeedeh wrote:

[Saeedeh]

Hi

I have installed libsvm from weka.

 weka version that I used is 3.7.11.

for installing libsvm:

when you open weka, in weka gui chooser => open tools menu

in tools menu select package manager. there are lots of library here.

you should choose libsvm.

after selecting libsvm, click install and then ok.

libsvm can install in this way easily.

you can also download other library in this way. 

[Bhoomin Pandya]

Hi Saeedeh,

Thanks I will try ti use 3.7.11 also when actual need arise. I could set up the LibSVM through the CLASSPATH in 3.6.11. But I am still not able to have LibLINEAR in the CLASSPATH.

What I gatherd from literature about the implementation of SMO (which originally handles two classes) is part of LibSVM. LibSVM need various parameters selected and that it requires proper parameter optimization.

I remembered something from Dr Ian Witten's Class Videos (don't remember exact which one) & I tried to check weather.nominal.arff show performance of 64.2857% with SMO & LibSVM (both with default setting). This may be just a coincidence as there are many parameters that could be different.

I tried to check another 2 class data i.e. german credit. With default settings the SMO preformance is 75.1% (with filter type = normalize training data, Linear Kernel = Linear Kernel: K(x,y) = <x,y>, takes 0.77 sec) where as LibSVM is 68.7% (1.18 sec) Please note, this also shows about 6% higher accuracy of SMO, similar to your dataset.

The same value of accuracy is achievable 75.2% with (normalize=TRUE, kernel type = linear: u'*v, takes 0.28 sec).

Thus, there is a difference in the default parameters setting for SMO & LibSVM that gives rise to difference in the performance. But once these parameters are properly set the efficiencies are comparable. Please note there are four Kernel Types (linear, ploynomial, radial, sigmoid) on both SMO and LibSVM along with various setting that need to be taken into account.

I hope that help you solve your problems.

Best Luck. Cheers
Bhoomin

On Sunday, June 8, 2014 4:13:23 PM UTC+5:30, Saeedeh wrote:

[Saeedeh]

I had a same problem in setting up class path for libsvm, so I prefered to use weka 3.7.11 to install it.

thanks for your comprehensive answer about smo and libsvm.

 I think I should study more about the difference of them. because libsvm use SMO-type, and I d'not really know the difference of them. if I  find more information about them, I will share it here.