Structure Learning for Groups of Variables in Nonlinear Time-Series Data with Location-Scale Noise

Genta Kikuchi, Shohei Shimizu
Proceedings of the 2023 Causal Analysis Workshop Series, PMLR 223:20-39, 2023.

Abstract

Learning causal structures from observational data has recently attracted considerable attention.Although many studies have focused on uncovering the connections between scalar random variables,estimation algorithms for groups of variables—particularly for multiple groups of variables—remain scarce.This paper proposes a novel differentiable algebraic constraint that can be used along with existing continuous optimization-based structure-learning algorithms to learn the causal relationships among groups of variables.Considering the complex functional relationships among variables in real-world scenarios, we propose a structure-learning algorithm for nonlinear time-series data with location-scale noise.Experimental results for synthetic and real-world data indicate that the proposed group acyclicity constraint significantly increases the estimation accuracy for the causal relationship among the groups of variables and verify the effectiveness of the proposed structure-learning algorithm.

Cite this Paper

BibTeX
@InProceedings{pmlr-v223-kikuchi23a, title = {Structure Learning for Groups of Variables in Nonlinear Time-Series Data with Location-Scale Noise}, author = {Kikuchi, Genta and Shimizu, Shohei}, booktitle = {Proceedings of the 2023 Causal Analysis Workshop Series}, pages = {20--39}, year = {2023}, editor = {Kummerfeld, Erich and Ma, Sisi and Rawls, Eric and Andrews, Bryan}, volume = {223}, series = {Proceedings of Machine Learning Research}, month = {14 Aug}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v223/kikuchi23a/kikuchi23a.pdf}, url = {https://proceedings.mlr.press/v223/kikuchi23a.html}, abstract = {Learning causal structures from observational data has recently attracted considerable attention.Although many studies have focused on uncovering the connections between scalar random variables,estimation algorithms for groups of variables—particularly for multiple groups of variables—remain scarce.This paper proposes a novel differentiable algebraic constraint that can be used along with existing continuous optimization-based structure-learning algorithms to learn the causal relationships among groups of variables.Considering the complex functional relationships among variables in real-world scenarios, we propose a structure-learning algorithm for nonlinear time-series data with location-scale noise.Experimental results for synthetic and real-world data indicate that the proposed group acyclicity constraint significantly increases the estimation accuracy for the causal relationship among the groups of variables and verify the effectiveness of the proposed structure-learning algorithm.} }
Endnote
%0 Conference Paper %T Structure Learning for Groups of Variables in Nonlinear Time-Series Data with Location-Scale Noise %A Genta Kikuchi %A Shohei Shimizu %B Proceedings of the 2023 Causal Analysis Workshop Series %C Proceedings of Machine Learning Research %D 2023 %E Erich Kummerfeld %E Sisi Ma %E Eric Rawls %E Bryan Andrews %F pmlr-v223-kikuchi23a %I PMLR %P 20--39 %U https://proceedings.mlr.press/v223/kikuchi23a.html %V 223 %X Learning causal structures from observational data has recently attracted considerable attention.Although many studies have focused on uncovering the connections between scalar random variables,estimation algorithms for groups of variables—particularly for multiple groups of variables—remain scarce.This paper proposes a novel differentiable algebraic constraint that can be used along with existing continuous optimization-based structure-learning algorithms to learn the causal relationships among groups of variables.Considering the complex functional relationships among variables in real-world scenarios, we propose a structure-learning algorithm for nonlinear time-series data with location-scale noise.Experimental results for synthetic and real-world data indicate that the proposed group acyclicity constraint significantly increases the estimation accuracy for the causal relationship among the groups of variables and verify the effectiveness of the proposed structure-learning algorithm.
APA
Kikuchi, G. & Shimizu, S.. (2023). Structure Learning for Groups of Variables in Nonlinear Time-Series Data with Location-Scale Noise. Proceedings of the 2023 Causal Analysis Workshop Series, in Proceedings of Machine Learning Research 223:20-39 Available from https://proceedings.mlr.press/v223/kikuchi23a.html.

Related Material