[2Win5] Poster session 2

[2Win5-13] Causal inference with latent variables extending likelihood-free overcomplete independent component analysis

〇Ryota Shishido¹, Masato Inoue¹ (1.Waseda University)

Keywords:Causal Inference, Independent Component Analysis

Causal inference is a method for estimating causal relationships from observed data, but accurate estimation is difficult when unobserved latent variables exist. Methods using overcomplete independent component analysis have been proposed, but are difficult to apply to multivariate data due to computational difficulties and assumptions on independent components. In this study, we extend the estimation method of likelihood-free overcomplete independent component analysis (2019) by C. Ding et al. and introduce the One-Latent-Component condition assumption (2023) by R. Cai et al. , which states that there is only one common causal latent variable for each observed variable. Validation on artificial data confirmed that the method converges to the correct parameters and outputs the graph correctly when the initial parameters are close to the correct ones. However, without assumptions on the initial parameters, the estimation was not correct. The devised method is effective when the causal relationship is known to some extent, and future work is needed to maintain performance with relaxed assumptions and to verify the method on other data.