Statistical assumption that class proportions differ between training and test data, but the relationship between features and labels remains constant.