Causal Discovery with Latent Confounders Based on Higher-Order Cumulants

Ruichu Cai,Zhiyi Huang,Wei Chen,Zhifeng Hao,Kun Zhang

2023-05-31

Abstract:Causal discovery with latent confounders is an important but challenging task in many scientific areas. Despite the success of some overcomplete independent component analysis (OICA) based methods in certain domains, they are computationally expensive and can easily get stuck into local optima. We notice that interestingly, by making use of higher-order cumulants, there exists a closed-form solution to OICA in specific cases, e.g., when the mixing procedure follows the One-Latent-Component structure. In light of the power of the closed-form solution to OICA corresponding to the One-Latent-Component structure, we formulate a way to estimate the mixing matrix using the higher-order cumulants, and further propose the testable One-Latent-Component condition to identify the latent variables and determine causal orders. By iteratively removing the share identified latent components, we successfully extend the results on the One-Latent-Component structure to the Multi-Latent-Component structure and finally provide a practical and asymptotically correct algorithm to learn the causal structure with latent variables. Experimental results illustrate the asymptotic correctness and effectiveness of the proposed method.

Machine Learning,Artificial Intelligence,Methodology

What problem does this paper attempt to address?

The paper is primarily dedicated to addressing the problem of causal discovery in the presence of latent confounders. Specifically, the research focuses on how to estimate the mixing matrix when there may be direct dependencies between observed variables and further determine the causal structure.

The paper proposes using higher-order cumulants to solve some computational challenges and local optima issues in Overcomplete Independent Component Analysis (OICA). The authors find that, in specific cases (e.g., when the mixing process follows a structure with only one latent component), a closed-form solution to the OICA problem can be obtained by utilizing higher-order cumulants.

Based on this finding, the contributions of the paper are as follows:
1. **Proposes a method based on higher-order cumulants** for estimating the mixing matrix, thereby identifying latent variables and determining the causal order.
2. **Defines the "one latent component structure"**, which is the case where two observed variables share one latent confounder, and provides a closed-form solution for estimating the mixing coefficients under this structure.
3. **Proposes an iterative algorithm** that extends the results from the "one latent component structure" to the "multiple latent component structure" by progressively removing identified shared latent components, thus providing a practically feasible and asymptotically correct causal structure learning algorithm.

Experimental results show that the proposed method is effective both theoretically and practically, capable of correctly recovering the causal structure. Additionally, the paper discusses the limitations of the method in real-world applications.