Connecting Data to Mechanisms with Meta Structual Causal Model

Gong Heyang

Keywords: meta-SCM, cyclic causal models, sufficient activated mechanisms

Abstract: Recent years have seen impressive progress in theoretical and algorithmic developments of causal inference across various disciplines in science and engineering. However, there is still some unresolved theoretical problems, especially for cyclic causal relationships. In this article, we propose a meta structure causal model (Meta-SCM) framework inspired by understanding causality as information transfer. A key feature of our framework is the introduction of the concept of \emph{active mechanisms} to connect data and the collection of underlying causal mechanisms. We show that the Meta-SCM provides a novel approach to address the theoretical complications for modeling cyclic causal relations. In addition, we propose a \emph{sufficient activated mechanisms} assumption, and explain its relationship with existing assumptions in causal representation learning. Finally, we conclude the main idea of the meta-SCM framework with an emphasis on its theoretical and conceptual novelty.

One-sentence Summary: A meta-SCM framework for cyclic causal modeling and the sufficient activated mechanisms assumption.