Subpopulation Shift
A structured special case of distribution shift in which the mixture proportions of latent subpopulations differ between training and test data, while the conditional distributions within subpopulations remain approximately the same.
Timeline (last 90 days)
Theoretical analysis shows that multiple configurations (including significantly imbalanced ones) can perform comparably to balanced configurations.
Subgroup shift is a structured special case of covariate shift, in which the conditional distribution p(y|x) is roughly the same within each subgroup, but the mixture proportions of the subgroups differ between training and testing.
Robustness under subgroup shift does not require the model to learn entirely new decision rules, but rather to achieve a more balanced error distribution across subgroups.
Adjusting the subgroup proportions in the training data may work, whereas simply increasing the amount of data may not solve the subgroup shift problem.
All Facts (4)
Theoretical analysis shows that multiple configurations (including significantly imbalanced ones) can perform comparably to balanced configurations.
50%UnverifiedSubgroup shift is a structured special case of covariate shift, in which the conditional distribution p(y|x) is roughly the same within each subgroup, but the mixture proportions of the subgroups differ between training and testing.
50%UnverifiedRobustness under subgroup shift does not require the model to learn entirely new decision rules, but rather to achieve a more balanced error distribution across subgroups.
50%UnverifiedAdjusting the subgroup proportions in the training data may work, whereas simply increasing the amount of data may not solve the subgroup shift problem.
50%