Spectrum is now read-only. Learn more about the decision in our official announcement.


Building and managing training datasets for machine learning


multi-class labeling

December 4, 2019 at 9:39pm

multi-class labeling

December 4, 2019 at 9:39pm
Hello, all! I am wondering if the LF allows the rules for defining multiple classes? for example: Rule-> class A and class B or Rule -> class A or Class B. Also, if LabelModel allows such LFs. if not, what's the best practice. Thanks!

December 4, 2019 at 11:54pm
Snorkel can support multi-class classification, as shown in this tutorial!
The rules can define multiple classes as in they can output abstain or one of the potential classes. We do not currently support an LF assigning multiple labels (AND, OR) but good to know there's interest here! The basic approach that would work is to treat it as a $k$-wise multi-label classification problem as $k$ separate classification tasks given there are no complex dependencies or constraints among the different classes.
We have a discussion around the multilabel LF case here. A similar topic around whether an LF can specify "NOT class A" is also discussed here.