Research map / Probabilistic and scientific modeling / Causal discovery

Integrating Background Knowledge for Scalable Causal Discovery

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Interventional causal discovery · 0 citations (OpenAlex)

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.

Extends constraint-based causal discovery (PC, SNAP, MB-by-MB, LDECC, LOAD) to use direct-relation background knowledge (adjacencies, orientations, gaps) during discovery, reducing CI tests while preserving soundness/completeness.

Problem

Causal discovery is computationally expensive in high dimensions, and expert background knowledge is typically applied only as post-processing rather than used to avoid CI tests during discovery.

Approach

Integrates background knowledge on adjacencies, orientations and gaps directly into the search of PC, SNAP, MB-by-MB, LDECC and Markov blanket algorithms; known gaps are handled by delayed search for separating sets rather than removed immediately, and candidate separating sets are restricted to possible parents.

Key results

• Background knowledge consistently reduces CI tests and running time for all algorithms, with improvements reaching an order of magnitude for PC-BK and LDECC+-BK.
• With oracle d-separation tests, all methods obtain more accurate causal effect estimates (lower intervention distance) as background knowledge increases.
• PC-BK and SNAP-BK benefit most from known gaps, while local methods benefit most from orientation knowledge; ARTH150 without BK was prohibitively slow for PC and LDECC+.

Surprising findings

• Naively removing known gaps without finding separating sets can lead to incorrect outputs even with consistent BK and oracle CI tests.
• SNAP(k)-BK becomes increasingly slow at higher BK rates under imperfect background knowledge.

Limitations

• Assumes no latent confounders or selection bias in main results; latent setting only preliminary.
• LDECC-BK is not complete and cannot propagate orientation background knowledge outside the Markov blanket.
• Imperfect background knowledge increases intervention distance and can slow SNAP(inf)-BK.

Experimental setup

Datasets: MAGIC-NIAB ECOLI70 ARTH150 ANDES

Benchmarks: MAGIC-NIAB ECOLI70 ARTH150 ANDES

Metrics: number of CI tests running time intervention distance

Baselines: PC SNAP MB-by-MB LDECC LOAD post-processing with background knowledge

Builds on: PC SNAP MB-by-MB LDECC LOAD

Closest papers by meaning

• Sound and Complete Causal Identification with Latent Variables Given Local Background Knowledge (NeurIPS 2022 · similarity 0.90)
• CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery (NeurIPS 2026 · similarity 0.90)
• THE ROBUSTNESS OF DIFFERENTIABLE CAUSAL DISCOVERY IN MISSPECIFIED SCENARIOS (ICLR 2025 · similarity 0.89)
• Adaptive Online Experimental Design for Causal Discovery (ICML 2024 · similarity 0.88)
• Causal Discovery with Fewer Conditional Independence Tests (ICML 2024 · similarity 0.88)
• A Versatile Causal Discovery Framework to Allow Causally-Related Hidden Variables (ICLR 2024 · similarity 0.88)