Are there any strategies to learn a model that can classify data from one domain using only data from a different domain for training?
For example, suppose I have a bunch of data from two different domains (Da and Db), and I want to train a model using data from those two domains that can then turn around and classify data from a third domain (Dc). Some of the features are shared between Da and Db, and some are not. Some of the features from Dc appear in Da and Db, but there are also some that do not. However, the labels in each domain are the same. So, I want to classify each instance as either class X or class Y, and I have labeled instances of class X and Y for Da and Db, and each instance I get from Dc will also either be of class X or class Y.
With the only information on Dc being which features are used and the classes to classify, is it even possible to determine which features specific to Dc are useful for classification, using the data and features from Da and Db? I know there are transfer learning techniques for cases where you have access to training data from Dc, labeled or no, but are there any techniques that can perform classification without any data from Dc? Intuitively I suspect not, because there is no way to know anything about the features specific to Dc without any training data from Dc. With some training data from Dc (labeled or not), it seems like current transfer learning techniques would be able to construct a mapping from features specific to Dc to features from Da or Db. If it is impossible as I suspect, are there ways to incrementally learn the mapping as each instance is introduced from Dc?