Someone just asked about the workings of class_weight on SO:
https://stackoverflow.com/questions/30972029/how-does-the-class-weight-parameter-in-scikit-learn-work/30982811?noredirect=1#comment99840069_30982811

Indeed I don't think that's clear from the glossary:
https://scikit-learn.org/dev/glossary.html#term-class-weight

Maybe we should add that it changes the loss-function in an algorithm specific way?
Or at least that it "usually" does that? I think RandomForests can or will be able to subsample as well.