I am performing a 12 class SVM classification with LIBSVM. My class labels are integers of {1,2,..., 12}. svmpredict sometimes returns an array of zeros instead of the predicted class labels. What would be the reason for this, and how do I prevent it?
Ohhh there are several reasons. The key two, I think, are:
1) the features ids of the test vector are not found in the training set, thus the scoring ends up with zero. This might be also due to some problem with your programming.
2) It is, in fact, zero, which is rare if you have many features and a balanced training set.
Please make sure that all the inputs of the functions are strictly corresponding to the usage of libsvm tool.
My software normalized all values before training because I was using LIBSVM mainly for SVR. When I tried SVC for a different problem it returned zeroes, too. Are you sure you use integer labels?
Edit: I read over "sometimes", shouldn't be this problem then.
There does not seem to be any problem with the input data and the labels. I wonder if there is a threshold that determines the reliability of the classifier model which returns zero if the threshold is not met.
- Badges
- Science topic