Distribution Shift

We focus on a challenging distribution-shift setting in which the causal and anticausal variables of the target are unobserved. Leaning on information theory, we develop feature selection and engineering techniques for the observed downstream variables that act as proxies. We identify model-stabilizing proxies and moreover utilize auxiliary training tasks to answer counterfactual questions that stabilize our models.

Importance weighting via probability density function estimation is unstable and requires parameter tuning. We demonstrate that reweighting according to cumulative distirbution functions is stable without parameter tuning.

We demonstrate that amplification processes lead to biases in TCGA data.