I am working on optimizing Gene selection in microarray data for Cancer Classification. I am going to use SVM in (libsvm) as wrapper approach to evaluate Gene subsets using 10 K fold cross validation.
Microarray data consider as huge dimensional data ( i.e Lymphoma data set consists 4026 Genes 'features' and 62 instances and 3 class labels).
Does libsvm support multiclass classification, As in my work, Lymphoma & MLL has 3 classes?
What is the appropriate svm type and kernal type and parameters for the chosen kernal (c,gamma, etc...) in LIBSVM multi class classification like microarray data?
Yes, you can classify multiple classes using LIBSVM. Your data set is sufficient to train and test the classifier. As for the kernal type, it is simply a hit and trail method, some time a linear kernal perform batter than gaussian and polynomial kernal.
To test kernal accuracy you can perform 10 fold cross validation using different kernals. Also, create ROC plot and calculate "AUC" to test the robustness of the models.
Similarly to choose C, gamma value is again a hit and trail method. These values depends on kernal functions. There are many other classifiers in weka which you can try to classify your data including LIBSVM.
which one of these svm type (C-SVC , nu-SVC ) used for multicast classification?
How about class label, assume that i have 3 classes, d i need to in addition of (scaling and format the values ) apply the same reprocessing OR keep the class values without changes, for instance :
3 1:4.3 2:55 3:89
2: 1:33 2 66 3:76
Finally, we pass the data as the above format to the SVM-Crossvalidation method in libsvm, May I right????
The types of classifier and kernel that are best depend on the data. Normally you would start with C-SVC and the radial basis kernel and try a grid of values for the penalty and gamma using cross-validation to decide the best parameter values. Then try alternative classifiers and kernels. With nu-SVC you have the extra nu parameter to optimise.
two papers should be refered to when you employ Libsvm:
LIBSVM: A Library for Support Vector Machines
A Practical Guide to Support Vector Classification.
You can google it to find the pdfs.
the famous and simple way for parameter optimization of svm is grid search, please refer to:
A Practical Guide to Support Vector Classification.
Some of my studies use meta-heuristic methods to optimize the parameters, please read the following paper for a brief overview about parameter optimization of svm:
A PSO and pattern search based memetic algorithm for SVMs parameters optimization http://www.sciencedirect.com/science/article/pii/S0925231213002038
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Bioinformatics
- Bioinformatic Tools