[1Win4] Poster session 1

[1Win4-14] Multivariate causal structure learning based on pairwise causal inference

〇Masaharu Hasuike¹, Izuru Miyazaki² (1.JTEKT Corporation, 2.TOYOTA CENTRAL R&D LABS., INC.)

Keywords:Causal discovery, non-additive noise

Discovering causality is crucial for the advancement of natural science and engineering. Recently, causal structure discovery has been actively studied. However, most existing methods assume additive noise, which leads to poor estimation accuracy in the presence of non-additive noise. In this paper, we propose a method for learning causal structures based on pairwise inference. Specifically, we introduce a method called graph-RECI, which utilizes Regression Error-based Causal Inference (RECI) as a pairwise inference approach. RECI has theoretical validity even in the presence of non-additive noise. We compared the proposed method with SCORE, a state-of-the-art (SOTA) approach, through numerical experiments using simulated data. For linear data, the proposed method outperformed SCORE except in cases with additive noise and weak causality. For nonlinear data, the proposed method was competitive with SCORE except in cases with additive noise and strong causality.