[Question] How to use Meka to train and evaluate a semi-supervised model?

[bundit786]

Hi Meka Team,

I have some experience using Meka to train and evaluate a supervised model. The semi-supervised model, which is stated to be a capable approach but with little information about how to use it. So, can you give me an example of how train and evaluate a model using a basic technique like pseudo-labelling?

Thank you.

[fracpete]

[from a quick look at the code...]

• Semi-supervised classifiers implement the meka.classifiers.multilabel.SemisupervisedClassifier interface, which has the method introduceUnlabelledData(Instances).
• The only classifiers implementing this interface are EM and CM (package meka.classifiers.multilabel.meta)
• The meka.classifiers.multilabel.Evaluation class (which is used when evaluating a classifier from the command-line or through the GUI) automatically supplies the test set (without its labels) to this method. Test set can either be a supplied one, one generated from a train/test split or the test part from a cross-validation fold pair.

[bundit786]

Thank you for getting back. I can now get CM in meka.classifiers.multilabel.meta using GUI. But for meka.classifiers.multilabel.SemisupervisedClassifier, is it available in GUI or only in command line?

[fracpete]

meka.classifiers.multilabel.SemisupervisedClassifier is just a Java interface for semi-supervised classifiers to implement, not an actual classifier that you can use. EM and CM implement this interface (and these are the only two semi-supervised classifiers in Meka).

[fracpete]

Assumed to be sorted.