"About SVM Regression using LibSVM in RapidMiner"

ikhwan

Dear all,

I have tried to do prediction using SVM. I use RBF kernel and epsilon-SVR, also I tried nu-SVR, as SVM type.
However, the results seem misleading. Most of predicted values are similar though they come from different attribute values.
In my dataset, I have three nominal attributes. I convert them, using "Nominal to Numerical", into numerical since SVM only support this attribute type. I also tried to convert them, using weka, into binary attributes, but there was no change to the final results.
Actually there are two other textual attributes, previoulsy I have included them as numeric one ( I convert into wordVector), but I still got the same results. 
Here is the detailed result: http://dl.dropbox.com/u/2902679/out_svm.pdf
And here is the process: http://dl.dropbox.com/u/2902679/process.pdf
Can anyone help me to fix this problem?
Another question: in "Set Role" node, there is option to set "target role". One of the possible value is "prediction". What is the purpose of this value? Should I use it for prediction? In this case, I only set it to "label".

Thanks in advance.

IngoRM

Hi,

since you are applying the model on the training data you should end up with perfect predictions after some parameter tuning. The most likely candidate: Increase the parameter "C" a lot and predictions will become closer to the actual values (at least if there isn't anything wrong with the data or the rest of the process setup).

Using the role "label" for defining which attribute should be predicted is perfectly fine. The role prediction can be used if the precition is for example derived with another tool (shame on you!  ) and you want to calculate, for example,  the performance with RapidMiner.

Cheers,
Ingo

ikhwan

Hi Ingo,

Thanks a lot for your hint. I finally use "optimize parameters" operator for finding the optimal SVM parameters. Now, I realized that this RapidMiner tool is really great.Previously I tried some other similar tools and found that RapidMiner is the one which has user friendly interface and many useful data mining features.

About the role "prediction", I think the name is kind of confusing especially for the newbie. I was thinking that role "label" is for classification and "prediction" is for numeric prediction, but now it is clear for me.

Cheers,
Ikhwan

IngoRM

Hi,

thanks for your kind words. I also believe that RapidMiner provides the best combination among all tools of huge analysis power, the possibility for designing complex processes, and - at least since version 5 - a cool and user friendly interface. But of course I am bit biased about that 


I think the roles of the attributes should now be better described in the new manual which is available since yesterday also in English:

http://rapid-i.com/rapidforum/index.php/topic,2432.0.html

Hope that helps. Cheers,
Ingo