Relational Structural Causal Models

This article introduces relational structural causal models, an extension of structural causal models that accommodates varying objects and their relationships, crucial for AI systems to reason about interventions and generalize to new combinations. It presents relational causal graphs and symbolic identification criteria to address the challenge of identifying causal and observational queries in the presence of unobserved confounding. Additionally, the proposed relational neural causal models demonstrate superior performance over non-relational baselines in simulated traffic scenarios, highlighting their potential for enhancing AI reasoning capabilities in dynamic environments.