TPAMI 2026 | 仅用两个变量破解混杂因素：CIC实现动力学因果推断与混杂变量重构

Core Viewpoint - The article discusses a novel framework for causal inference in complex biological systems, particularly under the influence of latent confounders, which are unobserved variables that can mislead causal relationships [2][4][25]. Group 1: Problem Statement - Accurately identifying causal relationships from observational time series data is a critical challenge across various fields, including life sciences and artificial intelligence [2]. - Latent confounders significantly complicate causal inference, as they cannot be directly measured and can create false causal associations [4][24]. - The traditional causal assumptions, such as the causal Markov assumption and faithfulness assumption, limit progress in causal inference within nonlinear dynamical systems [4][24]. Group 2: Proposed Methodology - The research introduces a new framework called Causal Inference under Latent Confounders (CIC), which aims to accurately identify causal directions and reconstruct unobserved confounders using only observed time series data [2][10][25]. - The CIC framework is based on the Takens delay embedding theorem, which allows for the reconstruction of the original system's state space from time series data [11][24]. - The methodology involves transforming original time series data into a delay embedding space and applying orthogonal decomposition to separate shared and private information between variables [12][23]. Group 3: Applications and Results - The CIC framework has been successfully applied to various nonlinear dynamical systems and complex biological systems, including gene regulatory networks and neuronal networks [16][24]. - Performance results indicate that CIC can effectively infer causal relationships and reconstruct latent confounders in different coupling scenarios [18][20]. - The framework demonstrates superior performance compared to existing methods in multiple benchmark systems and real biological data [24][25].