Partial Causal Structure Learning for Valid Selective Conformal Inference under Interventions

2026-03-02 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

This research introduces a novel approach for selective conformal prediction in interventional settings, specifically addressing scenarios where the invariance structure is unknown and must be learned from data. The work presents a contamination-robust conformal coverage theorem, quantifying coverage degradation due to misclassified "unaffected" calibration examples through a function $g(\delta,n)$ that depends on contamination fraction $\delta$ and calibration set size $n$. It also proposes a task-driven partial causal learning formulation to estimate binary descendant indicators $Z_{a,i}=\mathbf{1}\{i\in\mathrm{desc}(a)\}$ for selective calibration, rather than inferring a full causal graph. Algorithms for descendant discovery using perturbation intersection patterns and approximate distance-to-intervention estimation via local invariant causal prediction are provided. Experiments on synthetic linear structural equation models (SEMs) show that the corrected procedure maintains $\ge 0.95$ coverage even with up to $\delta=0.30$ contamination, where uncorrected selective conformal prediction drops to $0.867$. A proof-of-concept on Replogle K562 CRISPR interference (CRISPRi) data further demonstrates its real-world applicability.

Key takeaway

For AI Researchers developing robust uncertainty quantification methods in interventional settings, you should consider integrating partial causal structure learning to maintain valid selective conformal inference. This approach allows for tighter uncertainty sets by identifying exchangeable calibration examples, even when the underlying invariance structure is initially unknown. Implementing the proposed contamination-robust procedures can ensure reliable coverage guarantees, as demonstrated by maintaining $\ge 0.95$ coverage under significant contamination.

Key insights

Selective conformal prediction under interventions can be robustly maintained by learning partial causal structures.

Principles

• Exchangeability holds within intervention subsets.
• Contamination degrades coverage predictably.
• Partial causal learning suffices for selective calibration.

Method

The method involves a contamination-robust conformal coverage theorem, task-driven partial causal learning for descendant indicators, and algorithms for descendant discovery via perturbation intersection patterns and local invariant causal prediction.

In practice

• Apply to genomic perturbation experiments.
• Use for tighter uncertainty sets.
• Control contamination up to $\delta=0.30$.

Topics

• Selective Conformal Prediction
• Causal Structure Learning
• Interventional Settings
• Invariant Causal Prediction
• Genomic Perturbations

Best for: AI Researcher, AI Scientist, Research Scientist

Related on AIssential

• How Useful is Causal Invariance for Domain Adaptation in Finite-Sample Settings? — Causal invariance improves supervised domain adaptation in finite-sample settings when target-risk margins between candidate models are sufficiently large.
• Multi-Agent Conformal Prediction with Personalized Statistical Validity — PFWCP offers personalized, statistically valid uncertainty quantification in multi-agent settings, overcoming data heterogeneity and privacy limits.
• Causal Identification from Counterfactual Data: Completeness and Bounding Results — The CTFIDU+ algorithm enables counterfactual identification using newly characterized realizable Layer 3 data.
• Instrumental and Proximal Causal Inference with Gaussian Processes — DGP provides reliable epistemic uncertainty quantification for IV and Proxy causal inference, recovering kernel estimators.
• Correlation Is Not Enough: Embedding Human Metadata for Individual Causal Discovery — Pretrained biomedical language models struggle with cross-domain causal inference due to false embedding proximity, requiring specialized contrastive training.
• Unsupervised Causal Abstractions Discovery — This work proposes an unsupervised method to discover high-level causal abstractions directly from low-level observational data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.