Year: 2026
Type: article
Abstract: Digital Twins (DTs) are often engineered as modular models, but in real deployments they operate as interconnected subsystems whose variables influence each other. Composing DTs requires recovering cross-module dependencies interpretably and reusably. Causal discovery methods like the Independent Component Analysis-based Linear Non-Gaussian Acyclic Model (ICA-LiNGAM) are attractive for this, inferring directed acyclic graphs (DAGs) from observational data under linear, non-Gaussian assumptions; however, when extending an initially known causal structure with additional candidate variables, purely noisy or irrelevant signals can severely degrade LiNGAM’s separation step, producing spurious edges and unstable graphs. This paper proposes NA-LiNGAM (Noise-Aware LiNGAM), a LiNGAM-based algorithm applied to DT composition that screens candidate variables before integrating them into the causal graph. NA-LiNGAM augments ICA-LiNGAM with a graph validation score quantifying edge robustness through three statistical checks: (i) conditional regression significance, (ii) permutation-based significance, and (iii) bootstrap stability. Using this score, NA-LiNGAM selects new variables that improve graph fidelity while discarding noise. We also introduce a fast variant reducing search complexity from exponential to linear. We validate NA-LiNGAM on synthetic benchmarks and real-world datasets, using Area Under the Curve for Precision-Recall (AUC-PR), Structural Hamming Distance (SHD), and Structural Intervention Distance (SID). Results show NA-LiNGAM substantially reduces spurious edges and preserves the true causal structure under high-noise settings, yielding more stable compositions than baseline LiNGAM and competitive performance against state-of-the-art methods.
Institution: Universidad de Málaga
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,200
Awards: G1092424111, G1175115578, G1708521704, G2065685387, G2262748287, G24120701, G2601397997, G294895050, G300979063, G3480869486, G3681454997, G4042783231, G410571670, G4175246066, G425668148, G4502321410, G451917667, G5080475149, G5195188938, G523076202, G5539308778, G5967599077, G6033112984, G7084143925, G890014953
OpenAlex ID: W7204094905