Learning Coarsened Causal DAG Models

Alex Markham, Pratik Misra

Keywords: graphical causal abstraction

TL;DR: we propose a notion of graphical causal abstraction, which frames the space of abstractions as a lattice, which is conducive to search algorithms

Abstract: Graphical models are a powerful tool for representing causal relationships among jointly distributed random variables, especially concerning data from across different experimental settings. However, it is not always practical or desirable to estimate a causal model at the granularity of features in a particular dataset. There is a growing body of research on causal abstraction to address such problems. We provide graphical identifiability results and propose an algorithm for directly and efficiently learning abstract causal graphs from data, as well as theoretical insights about the lattice structure of this search space. As proof of concept, we apply our algorithm to synthetic data as well as a real dataset containing measurements from protein-signaling networks. *This is a work-in-progress.*

Submission Number: 12