Invariant Causal Prediction for Nonlinear Models

摘要: An important problem in many domains is to predict how a system will respond interventions. This task inherently linked estimating the system's underlying causal structure. To this end, Invariant Causal Prediction (ICP) (Peters et al., 2016) has been proposed which learns model exploiting invariance of relations using data from different environments. When considering linear models, implementation ICP relatively straightforward. However, nonlinear case more challenging due difficulty performing nonparametric tests for conditional independence. In work, we present and evaluate an array methods versions learning parents given target variables. We find that approach first fits with pooled over all environments then differences between residual distributions across quite robust large variety simulation settings. call procedure "invariant distribution test". general, observe performance approaches critically dependent on true (unknown) structure it becomes achieve high power if parental set includes than two As real-world example, consider fertility rate modelling central world population projections. explore predicting effect hypothetical interventions accepted models ICP. The results reaffirm previously observed role child mortality rates.