Research map / Probabilistic and scientific modeling / Causal discovery
Local Causal Discovery Without Causal Sufficiency
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 LocalMAG and the LatentLCD algorithm to recover local causal structure (parents, children, and latent-confounded bidirectional neighbors) of a target variable without assuming causal sufficiency, evaluated on synthetic random DAGs and the BARLEY network.
Problem
Local causal discovery around a target variable when latent common causes exist, which violates causal sufficiency and causes traditional local algorithms to misidentify direct effects of latent variables as parents or children.
Approach
Defines a local maximal ancestral graph (LocalMAG) including parents, children, and effects of latent causes (bidirectional neighbors) of the target, and learns it via MMPC skeleton discovery, collider tests for V-structures, and a FindBI procedure that identifies bidirectional edges by checking whether both target and neighbor are colliders.
Key results
- On 25-node synthetic graphs at 3000 samples, LatentLCD improves orientation accuracy by 27%, 13%, and 38% over FCI, RFCI, and ICD respectively.
- On the BARLEY network with DG25 target at sample size 1500, LatentLCD increases F1 by 44%, 56%, and 35% over PCD-by-PCD, CMB, and PSL.
- LatentLCD achieves the highest skeleton F1 and recall across all sample sizes on synthetic 25- and 35-node graphs.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LatentLCD instead of RFCI | Global causal discovery · 25 nodes, 3000 samples | OAcc: 0.45±0.09 -> 0.58±0.09 | Helped |
| Use LatentLCD instead of RFCI | Global causal discovery · 25 nodes, 3000 samples | ske-F1: 0.89±0.07 -> 0.89±0.06 | No clear effect |
| Use LatentLCD instead of ICD | Global causal discovery · 25 nodes, 3000 samples | CItests: 456±96 -> 1022±310 | Hurt |
| Use LatentLCD instead of RFCI | Global causal discovery · 35 nodes, 3000 samples | OAcc: 0.46±0.09 -> 0.60±0.08 | Helped |
| Use LatentLCD instead of RFCI | Global causal discovery · 35 nodes, 3000 samples | ske-F1: 0.89±0.06 -> 0.89±0.05 | No clear effect |
| Use LatentLCD instead of ICD | Global causal discovery · 35 nodes, 3000 samples | CItests: 894±166 -> 1671±424 | Hurt |
| Use LatentLCD instead of PCD-by-PCD | Local causal discovery, target FRSPDGA · BARLEY, 3000 samples | F1: +30% | Helped |
| Use LatentLCD instead of CMB | Local causal discovery, target FRSPDGA · BARLEY, 3000 samples | F1: +30% | Helped |
| Use LatentLCD instead of PSL | Local causal discovery, target FRSPDGA · BARLEY, 3000 samples | F1: +15% | Helped |
| Use LatentLCD instead of PCD-by-PCD | Local causal discovery, target DG25 · BARLEY, 1500 samples | F1: +44% | Helped |
| Use LatentLCD instead of CMB | Local causal discovery, target DG25 · BARLEY, 1500 samples | F1: +56% | Helped |
| Use LatentLCD instead of PSL | Local causal discovery, target DG25 · BARLEY, 1500 samples | F1: +35% | Helped |
Limitations
- Evaluation limited to synthetic random DAGs and one benchmark network with a single latent variable
- Relies on causal faithfulness and reliable CI tests
- Current work is local only; extending to a local-to-global algorithm is future work
Experimental setup
Datasets: synthetic random DAGs with latent common causes BARLEY Bayesian network
Benchmarks: BARLEY Bayesian network (DG25, FRSPDAG, DGV1059 targets)
Metrics: orientation accuracy skeleton F1 skeleton recall F1 precision recall number of CI tests
Baselines: FCI RFCI ICD PCD-by-PCD CMB PSL
Builds on: PC/MMPC FCI RFCI PSL Zhang 2008 orientation rules
Closest papers by meaning
- Local Identifying Causal Relations in the Presence of Latent Variables (ICML 2025 · similarity 0.91)
- Local Causal Structure Learning in the Presence of Latent Variables (ICML 2024 · similarity 0.89)
- A Versatile Causal Discovery Framework to Allow Causally-Related Hidden Variables (ICLR 2024 · similarity 0.88)
- Hybrid Top-Down Global Causal Discovery with Local Search for Linear and Nonlinear Additive Noise Models (NeurIPS 2024 · similarity 0.88)
- Identification of Causal Structure with Latent Variables Based on Higher Order Cumulants (AAAI 2024 · similarity 0.87)
- Differentiable Causal Discovery for Latent Hierarchical Causal Models (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.