Watch the recording here.

Zoom link.

Date: May 1

Time: 2 p.m.

Location: KEC 1001

Abstract: Causal graphs are a popular framework for guiding and automating causal reasoning. However, graph-based causal reasoning is often computationally and algorithmically challenging, particularly given the diversity of causal reasoning tasks and graph types studied in the literature. A principled way to manage this diversity is through the use of algorithmic primitives: well-understood and optimized subroutines that can be reused across multiple problems. We analyze and compare three core primitives in graph-based causal reasoning: reachability, moralization, and latent projection. We introduce CIfly, a framework that isolates modular variants of graph reachability as a reusable core operation underlying a broad class of causal reasoning algorithms. At the center of CIfly is a rule-table schema that allows succinct specification of any CIfly algorithms. We show that CIfly algorithms run in linear time and can solve many important causal reasoning tasks. In contrast, we prove that moralization and latent projection are computationally equivalent to Boolean matrix multiplication, and therefore more computationally expensive than is commonly stated in the literature or than CIfly algorithms.  Our results position reachability in general and CIfly in particular as a central tool for scalable causal reasoning.

Bio: Leonard Henckel is an assistant professor at University College Dublin. His research focuses on how to estimate causal effects from observational data using graphical models. Henckel is particularly interested in instrumental variable based approaches, tools for efficient causal effect estimation and algorithmics for causal inference. Previously, he was a postdoctoral researcher supervised by Jonas Peters at the University of Copenhagen and obtained his PhD supervised by Marloes H. Maathuis at the Seminar for Statistics, ETH Zurich.