CAUSALITY FROM A DISTRIBUTIONAL ROBUSTNESS POINT OF VIEW

By Nicolai Meinshausen. 2018

88 citations recorded by the source corpus.

We show some connections between distributional robust and causal inference. Distributional robustness is here understood as optimizing the predictive accuracy for a whole class of distributions instead of just a single target distribution. The class of distributions considered is often a ball around the empirical distribution, using a Wasserstein or Kullback-Leibler distance. Causal inference can be seen as a special case of distributional robustness, where the class of distributions is generated by a causal model. Different choices for these classes are discussed, which correspond to different assumptions about the type of interventions and presence or absence of latent confounding variables.

Subfields: Artificial Intelligence, Economics and Econometrics, Statistics and Probability, Causal Inference & Discovery, Statistical Learning Theory

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Economics and Econometrics

In Statistics and Probability

In Causal Inference & Discovery

In Statistical Learning Theory

Citations in the corpus

Toward Causal Representation Learning (2021, methodology): “To formalize this general intuition, one can consider a form of out-of-distribution generalization, which can be optimized by minimizing the empirical risk over a class of distributions induced by a causal model of the data [5], [169], [187], [204], [220].”
Towards Causal Representation Learning (2021, background): “induced by a causal model of the data [5, 204, 169, 189, 218].”
Invariant Risk Minimization (2019, background): “Some works in machine learning [45, 18, 21, 26, 36, 43, 34, 7] pursue similar questions.”
Conditional variance penalties and domain shift robustness (2017, background): “Causal inference can be seen to be a specific instance of distributional robustness, where we take F to be the class of all distributions generated under do-interventions on X (Meinshausen, 2018; Rothenhäusler et al., 2018).”
Preventing Failures Due to Dataset Shift: Learning Predictive Models That Transport (2018, background): “In safety-critical applications where preventing failures is crucial, we require unbounded invariance to perturbations which motivates the use of causal-based methods (Meinshausen, 2018).”
Towards Out-Of-Distribution Generalization: A Survey (2021, background): “…P ( P tr ) is often characterized by f -divergence or Wasserstein distance, different choices of P ( P tr ) will render DRO equivalent to causal inference in the structural equation model (SEM) context [159], which shows the inherent relationship between causality-based methods and DRO methods.”
Guarding Against Adversarial Domain Shifts with Counterfactual Regularization (2017, background): “anteed performance bound. Causal inference can be seen to be a specic instance of distributional robustness, where we take Fto be the class of all distributions generated under do-interventions on X(Meinshausen, 2018; Rothenhausler et al., 2018). Causal models thus have the dening advantage that the”
BREEDS: Benchmarks for Subpopulation Shift (2020, background): “Distributional robustness has also been connected to causality [Mei18]: here, the typical approach is to inject spurious correlations into the dataset, and assess to what extent models rely on them for their predictions [HM17; Arj+19; Sag+20].”