i am using follwoing code to train the classifier: model = svmtrain(train_label, train_data, '-s 0 -t 2'); is it fine for multiclass classification
for testing i am using following
[predict_label, accuracy, dec_values] = svmpredict(test_label, test_data, model);
is it enough?
There is no set-in-stone requirement/rule on how you want to train and test your model for multiclass classification using SVM (or any type of classification). It depends on the type of data you have and what you want to get out of it.
In most cases, you can use any combination of parameters to get results, which applies to your case as well. However, the accuracy/quality of your result depends on your approach to classifying the data that you have.
I can see that you are using an RBF kernel. You can use any kernel you want but you should be able to answer why you are using it. Sometimes a linear kernel works just as good or even better as compared to other kernels but it depends on your data.
You can ask yourself a few questions and examine your dataset to see if you are approaching the right direction. Have you visualized your data? If so, are the class separations linear or non-linear? How many features do you have? Do you need to reduce the dimensionality of the features? Is your dataset suffering from high bias/variance? Are you overfitting your data? These are only general questions since I do not know what specifically caters to your dataset.
feature vector dimension is 1x1876 and if i have 1000 instances i would be 1000x1876.........should i reduce feature vector dimensions?
I cannot exactly say if you should apply dimensionality reduction on your feature set without looking at it but that looks like a lot of features! If I am not mistaken, you are saying that you have 1876 features and 1000 instances of each feature, right? If that's the case, then you should check if you are overfitting your model (high variance) due to a relatively small number of samples as compared to the number of features. I would suggest that you get a lot more training samples. If that's not possible, then you can see if an appropriate dimensionality reduction technique does any good.
Just wanted to add something: Look into cross-validation as well if you can gather more training smaples.
Yes, that's enough.
For details of running multi-class SVM just check 'readme' file inside LIBSVM or read the following paper:
LIBSVM: A Library for Support Vector Machines
Article LIBSVM: A library for support vector machines
Thank you.......Your answers are very helpful for me....for now i am unable to find any option to plot training and testing features and support vectors in LIBSVM for multi class and multi-dimensional features. As i want to visualize features and boundery lines
This may sound silly but the new MS Excel is a good way to start viaualizing your training samples. You don't have to code anything, just copy and paste your data, select all or whatever portion you'd like to visualize and select line graph. Our eyes are the best tool for an initial screening and getting an idea of how the decision boundaries may look like (linear, non-linear, clustered etc.).
I'm not sure but I do not think LibSVM would have like a built-in function to plot and visualize features. You may have to code it yourself.
i find a page very helpful in plotting decision boudnries
https://sites.google.com/site/kittipat/libsvm_matlab