How does WEKA handle nominal attribute values without instances
I'm trying to fuse together two machine learning systems, one supervised, the other unsupervised. Both are multi-class systems and so far they've both been predicting the same range of classes (a nominal class attribute).
However, the actual problem requires more classes to be added for which I don't have training data. This means the supervised part of the system will never predict these, while the unsupervised part can handle it perfectly well (albeit with lower accuracy).
As a first step I have changed the header of my nominal class attribute to include all possible classes, even those for which I don't have training data. My question is: how are supervised classifiers from WEKA actually handling such untrained classes? What would the resulting likelihood distribution look like? Would it contain zero's for unchanged classes? Or chance-level values for them? I think the latter would be ideal for my purposes...