Relaxing Faithfulness with Intervention-Only Causal Discovery

2026-07-14 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

A new causal discovery framework, "Relaxing Faithfulness with Intervention-Only Causal Discovery," addresses limitations of traditional algorithms that rely on the "faithfulness" assumption. This assumption, requiring statistical dependence for causally linked variables, is often violated by "path cancellation" in natural systems, leading to incorrect causal structure identification. The authors introduce "intervention immediacy (II) faithfulness," a significantly milder assumption that tolerates such cancellations. This framework re-prioritizes "hard interventions" as the primary carriers of causal information, moving away from conditional independence (CI) testing. Two algorithms, Unrestricted Intervention Cardinality (UIC) and k-Restricted Intervention Cardinality (k-RIC), are proposed. UIC identifies the true graph using O(n^2) interventions of cardinality up to n-1. k-RIC identifies the graph up to its k-robust transitive closure when intervention cardinality is limited. Empirical validation on 10-node synthetic DAGs shows k-RIC, particularly with k=2 and CI-refinement, achieves an average Structural Hamming Distance (SHD) of 4-6, outperforming CI-based baselines like PC->Interv and IGSP by approximately half.

Key takeaway

For Research Scientists and Machine Learning Engineers designing causal discovery experiments, especially in fields like cell biology where hard interventions are feasible, you should prioritize intervention-based methods over traditional conditional independence tests. This approach, leveraging "intervention immediacy faithfulness," robustly identifies causal structures even in systems with path cancellation. Implement multi-node interventions (e.g., k=2) and a CI-refinement step to achieve significantly higher accuracy, reducing structural Hamming distance by approximately half compared to CI-first baselines.

Key insights

Intervention-only causal discovery, using "intervention immediacy faithfulness," robustly identifies causal structures by prioritizing hard interventions over conditional independence tests.

Principles

• Path cancellation violates standard faithfulness assumptions.
• Hard interventions reveal causal links directly.
• II faithfulness is parametrically milder than CI faithfulness.

Method

Algorithms perform single-node interventions to build a transitive closure. Edges are then pruned by testing if endpoints are conditional Most Immediate Children (MICs) given other parents or minimum vertex cuts, using hard interventions.

In practice

• Prioritize hard interventions for causal structure learning.
• Employ multi-node interventions for better accuracy.
• Refine intervention-derived graphs with CI tests.

Topics

• Causal Discovery
• Hard Interventions
• Faithfulness Assumption
• Path Cancellation
• Structural Causal Models
• k-RIC Algorithm

Code references

• honeybijan/Intervention-Only-Causal-Discovery

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Partial Causal Structure Learning for Valid Selective Conformal Inference under Interventions — Selective conformal prediction under interventions can be robustly maintained by learning partial causal structures.
• Meta-Dependence in Conditional Independence Testing — CIMD quantifies shared information between conditional independence tests, revealing how their violations co-occur under finite-sample uncertainty.
• Optimal Experiments for Partial Causal Effect Identification — Optimal experimental design for partially identifiable causal queries is NP-hard but can be significantly pruned using graphical criteria.
• Unsupervised Causal Abstractions Discovery — This work proposes an unsupervised method to discover high-level causal abstractions directly from low-level observational data.
• Proximal Mediation Analysis with Hidden Recanting Witnesses — Proximal causal inference identifies path-specific effects despite hidden recanting witnesses, enabling robust mediation analysis.
• A Survey on Federated Causal Discovery and Inference — Federated Causal Discovery and Inference are complementary stages in a unified pipeline for privacy-preserving causal reasoning.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.