Is that possible that we use bagging on the combination of a decision tree and a linear separator?
I know bagging could be used for the decision tree or linear separator individually. but how about the combination?
This is a rather complicated question.
Although bagging can be used with decision trees, the original approach doesn't work very well -- the subsampled populations produce highly correlated trees. The solution generally takes you in the direction of random forests, which are a variation on bagging.
Bagging itself is a form of model voting. Yes, you can have different models vote for the final result. You can even have a heterogeneous set of models developed independently to do the voting. Whether this produces an improved model, I don't know. Bootstrap aggregation (the formal name of bagging) is usually applied to one type of model -- although in a modified form for decision trees.