NeurIPS 2020

Active Invariant Causal Prediction: Experiment Selection through Stability


Meta Review

The authors propose an active causal learning framework that extends invariant causal prediction (ICP) to the active online setting. ICP assumes that the conditional distribution of the response, given its direct causes, remains invariant when intervening on arbitrary variables in the system. Several active learning policies for performing interventions and recovering the parent set of the target variable are also proposed. Overall the reviewers liked the paper but they have some reservations about the computational complexity of the proposed approach.