Original title: Block-Coordinate Descent Algorithm for Interventional Data in Directed Graphical Models
Authors: Wu, Jun
Document type: Papers
Conference/Event: Workshop on Uncertainty Processing - WUPES 2025 /13./, Třešť (CZ), 20250604
Year: 2025
Language: eng
Abstract: Computing maximum likelihood estimates in linear structural equation models is generally a difficult problem. The critical equations are usually non-linear and have numerous solutions, even for purely observational data. The block-coordinate descent (BCD) algorithm proposed by Drton et al. (2019)[1] is an efficient way to solve the optimization problem by decomposing it into a series of sub-problems with closed-form solutions, and which works with observational data. In this work, we describe the general problem of a BCD-type scheme for computing maximum likelihood estimates in linear structural equation models without hidden variables, integrating multiple observational and interventional environments. With interventional data, the degrees of both the original likelihood equations and the block-coordinate update equations could increase greatly. We study special setups in which the block optimization subproblems have a degree of at most 2 and provide closed-form solutions in these cases. Additionally, we discuss the potential applications of the model and algorithm to health and well-being data.\n
Project no.: EH22_008/0004583
Funding provider: GA MŠk
Host item entry: Proceedings of the 13th Workshop on Uncertainty Processing (WUPES’25), ISBN 978-80-7378-525-3

Institution: Institute of Computer Science AS ČR (web)
Document availability information: Fulltext is available at the institute of the Academy of Sciences.
Original record: https://hdl.handle.net/11104/0367844

Permalink: http://www.nusl.cz/ntk/nusl-684948


The record appears in these collections:
Research > Institutes ASCR > Institute of Computer Science
Conference materials > Papers