Key points are not available for this paper at this time.
Standard machine learning methods suffer distribution shifts between training and test data. This is an extended abstract of my talk on distribution shift adaptation.
Masashi Sugiyama (Mon,) studied this question.