I am using Libsvm(3.18) software with MATLAB interface to classify my data which consist of 4 classes. I am using -b 1 option in svmtrain and svmpredict functions.
In the output there are 4 prob.estimate values for each test sample. May be for 4 classes 4 values.According to me highest probability is the predict label class. But, unfortunately i am not finding this match.
Can anyone identify my mistake?May be i am wrong interpreting the matter?
If you want to classify new data (X) you should use the 'predict' function on a trained SVM model:
http://www.mathworks.com/help/stats/compactclassificationsvm.predict.html
label = predict(SVMModel,X) returns a vector of predicted class labels for predictor data X, based on the full or compact, trained SVM classifier SVMModel.
Thanks Emilson for the Answer.
I have tried your suggestion,
I have train my training data set using svmtrain function and stored it into my_model. Then I am writing predict(my_model,testing_file).My testing file contains features without their labels.As i want to predict the labels But, it is giving error 'Undefined function or method 'predict' for input arguments of type 'struct''.
Help on predict shows that ,'predict is meaningful only for time-domain data'
Can you focus on this problem?
Thanks.
The probability estimate you must be getting using LIbSVM must be correct. Though, your interpretation could be wrong. SVM will give highest probability to the class which it has recognized, not what it was supposed to be. So probably in your case, it could be a case of class misidentification by your model. Still, I didn't understood your problem very well. A bit more elaboration will probably help to solve this issue
Make sure you are using the 'predict' function that is in the Statistics and Machine Learning Toolbox (there are overloaded predict functions in Matlab).
Does your 'testing_file' contains the label information of test data set ?
No. I contains only binary values. In the original file information is there. But, during training-testing I have removed the information to convert my files in to binary.
The libsvm probability outputs appear in the order the class labels are ordered in your training data. For example if the class label 3 comes first in your training data, that is the first probability you get instead of the probability of the class label 1.
- Badges
- Science topic