1. how many data instances do you have? and the dimention of data? is the data too big? Maybe you can use some methods to select the informative data subsets for learning.

2. which kind parameter tuning do you use. If grid search, is the parameters range too big? Besides, With improper parameters, it may need a lot of time to train the model.

Thank you for your reply.

Regarding the no. of instances and dimension (colon-cancer: 2000 features, 62 instances) , (Gisette 5000 features, 6000 instances), (Leukemia: 7130 features, 38 instances). and I am using classification for feature selection.

Regarding parameter tuning, there is no such option (grid search) in LibSVM.

Your help is really appreciated.

so many features for your data... And so few instances for two of the datasets. You should do some feature selection procedure to reduce the feature sets.

What do you mean "using classification for feature selection"? Many feature selection methods can be applied (filter, wrapper, hybrid), among them, filter method (such as by using Relief-F, MI, corr,) is relatively efficient.

kernel function and its parameter, and C in SVM should be tuned. RBF kernel is also applied. but it also depends on the data. Improper setting will results in long-time consuming.

The reason why SVM is running slow is due to the Numbers of support vector which are running are more this take lot of running time you can reduce this by adopting various algorithm like RSVM to obtain the best result  by reducing the numbers of Support vector.

Hi, maybe you can use InfoGainAttributeEval in Weka to select the most relevant features.