Research map / Probabilistic and scientific modeling / Causal discovery

Local Identifying Causal Relations in the Presence of Latent Variables

ICML 2025 · Probabilistic and scientific modeling › Causal discovery · approach: Interventional causal discovery

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes LocICR, a fully local algorithm that identifies whether one variable is an invariant non-ancestor, invariant ancestor, or possible ancestor of another under latent confounding, without learning the full PAG.

Problem

Determining causal relationships between a specific pair of observed variables from observational data when latent variables may be present; global PAG-based methods are computationally expensive and redundant when only one causal relation is of interest.

Approach

Gives sound and complete local graphical characterizations (via augmented parent sets, augmented undirected neighbor sets, and Markov blanket local structure learning) of invariant non-ancestor, explicit/implicit invariant ancestor and possible ancestor relations, so that a pair's causal relation can be decided by conditional independence tests on local sets without learning the global PAG.

Key results

• LocICR achieves higher weighted F1 than all compared methods across all four benchmark networks and sample sizes from 1000 to 5000.
• LocICR requires substantially fewer conditional independence tests than global methods such as PC-ITC, PC-IDA, RFCI-LVIDA, RFCI-Zhang and ICD-Zhang.
• On real data, LocICR's findings on the General Social Survey and gene expression datasets align with domain knowledge in Duncan et al. (1972) and Wille et al. (2004).

Measured effects reported in the paper

Limitations

• Assumes oracle conditional independence tests for theoretical completeness; experiments rely on statistical tests with finite samples.
• Experiments use linear Gaussian SCMs with a fixed number of latent variables.
• Worst-case complexity depends on the number of maximal cliques of possible children and neighbors.

Experimental setup

Datasets: MILDEW ALARM WIN95PTS ANDES General Social Survey data Arabidopsis thaliana gene expression data (Wille et al. 2004)

Benchmarks: MILDEW ALARM WIN95PTS ANDES General Social Survey Arabidopsis gene expression (Wille et al. 2004)

Metrics: weighted precision weighted recall weighted F1 number of conditional independence tests (nTest)

Baselines: PC-ITC PC-IDA Local-ITC RFCI-LVIDA RFCI-Zhang M3HC-Zhang ICD-Zhang VIDA

Builds on: FCI RFCI MMB-by-MMB Local-ITC Fang et al. 2022 local causal discovery Total Conditioning Markov blanket discovery

Closest papers by meaning

• Local Causal Structure Learning in the Presence of Latent Variables (ICML 2024 · similarity 0.94)
• Differentiable Causal Discovery for Latent Hierarchical Causal Models (ICLR 2025 · similarity 0.91)
• Local Causal Discovery Without Causal Sufficiency (AAAI 2025 · similarity 0.91)
• Identification of Latent Confounders via Investigating the Tensor Ranks of the Nonlinear Observations (ICML 2025 · similarity 0.90)
• Identification of Causal Structure with Latent Variables Based on Higher Order Cumulants (AAAI 2024 · similarity 0.90)
• Identifiable Latent Polynomial Causal Models through the Lens of Change (ICLR 2024 · similarity 0.90)