[2D5-GS-2] Machine learning: Statistical learning

[2D5-GS-2-04] Adversarial Nuisance for Causal Effect Estimation

〇Akria Tanimoto¹ (1. NEC)

Keywords:Causal inference, Pessimistic inference

Causal inference is the problem of inferring the outcome of decisions that may differ from those taken in the past. One of the basic approaches is importance sampling, in which sample weights are generally the inverse of the propensity score, a nuisance parameter estimated a priori and plugged-in. A problem here is the estimated variance especially when the propensity score can be extremely small. This is particularly problematic when the randomness of the past decision is low or when the hypothesis space is complex, such as with deep models or high-dimensional data. We propose an end-to-end estimation method based on a generalized error upper bound by taking the worst-case for the uncertainty in the nuisance estimation.