Higher Order Cumulants-Based Method for Direct and Efficient Causal Discovery
Chen, Wei Peng, Linjun Huang, Zhiyi Cai, Ruichu Hao, Zhifeng Zhang, Kun
Chen, Wei
Peng, Linjun
Huang, Zhiyi
Cai, Ruichu
Hao, Zhifeng
Zhang, Kun
Supervisor
Department
Machine Learning
Embargo End Date
Type
Journal article
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Causal discovery plays a pivotal role in scientific inquiry and subsequent applications in prediction or decision-making. While many methods have been proposed, many of them rely on independence tests. However, these tests are difficult to implement and computationally intensive. In this article, we aim to propose a direct and computationally efficient method to determine the causal relationship between two observed variables in the linear non-Gaussian case. Building on the insight that cumulants provide information about the shape of a probability distribution, we show that interestingly, the (in)dependence between two observed variables can be directly inferred from the difference in the product of certain joint cumulants of these variables. This concept is named the cause difference criterion. Based on this criterion, we introduce two practical methods, high-order cumulant (HC) and HC-linear non-Gaussian acyclic model (LiNGAM), for causal discovery in the high-dimensional case. Theoretical analyses ensure the identifiability of the proposed criteria and methods. Experimental results indicate that our methods outperform most existing methods.
Citation
W. Chen, L. Peng, Z. Huang, R. Cai, Z. Hao and K. Zhang, "Higher Order Cumulants-Based Method for Direct and Efficient Causal Discovery," in IEEE Transactions on Neural Networks and Learning Systems, doi: 10.1109/TNNLS.2025.3622148
Source
IEEE Transactions on Neural Networks and Learning Systems
Conference
Keywords
Causal discovery, cumulants, identifiability, linear non-Gaussian causal model
Subjects
Source
Publisher
IEEE