[ad_1]
*=Equal Contributors
This paper was accepted on the workshop “Reliable Machine Studying for Healthcare Workshop” on the convention ICLR 2023.
When analyzing robustness of predictive fashions underneath distribution shift, many works concentrate on tackling generalization within the presence of spurious correlations. On this case, one sometimes makes use of covariates or atmosphere indicators to implement independencies in discovered fashions to ensure generalization underneath numerous distribution shifts. On this work, we analyze a category of distribution shifts, the place such independencies aren’t fascinating, as there’s a causal affiliation between covariates and outcomes of curiosity. This case is frequent within the well being house the place covariates may be causally, versus spuriously, associated to outcomes of curiosity. We formalize this setting and relate it to frequent distribution shift settings from the literature. We theoretically present why customary supervised studying and invariant studying won’t yield strong predictors on this case, whereas together with the causal covariates into the prediction mannequin can recuperate robustness. We reveal our theoretical findings in experiments on each artificial and actual knowledge.
[ad_2]