[Paper Review] Ancestral Causal Inference

Sara Magliacane, Tom Claassen+1 more|arXiv (Cornell University)|Jun 22, 2016

Bayesian Modeling and Causal Inference27 references16 citations

TL;DR

This paper introduces Ancestral Causal Inference (ACI), a logic-based method that improves scalability and reliability in constraint-based causal discovery by reasoning over ancestral (indirect) causal relations instead of direct ones. It achieves state-of-the-art accuracy with orders-of-magnitude speedup and provides confidence scores for predictions, enabling robust handling of observational and interventional data with background knowledge integration.

ABSTRACT

Constraint-based causal discovery from limited data is a notoriously difficult challenge due to the many borderline independence test decisions. Several approaches to improve the reliability of the predictions by exploiting redundancy in the independence information have been proposed recently. Though promising, existing approaches can still be greatly improved in terms of accuracy and scalability. We present a novel method that reduces the combinatorial explosion of the search space by using a more coarse-grained representation of causal information, drastically reducing computation time. Additionally, we propose a method to score causal predictions based on their confidence. Crucially, our implementation also allows one to easily combine observational and interventional data and to incorporate various types of available background knowledge. We prove soundness and asymptotic consistency of our method and demonstrate that it can outperform the state-of-the-art on synthetic data, achieving a speedup of several orders of magnitude. We illustrate its practical feasibility by applying it on a challenging protein data set.

Motivation & Objective

• To address the scalability and reliability issues in constraint-based causal discovery due to combinatorial explosion and unreliable independence tests.
• To improve accuracy and efficiency by shifting from direct to ancestral causal relation representation.
• To provide confidence scores for causal predictions to enable ranking and reliability assessment.
• To integrate observational and interventional data and background knowledge seamlessly in a logic-based framework.
• To demonstrate practical feasibility on a challenging real-world protein dataset with faithfulness violations.

Proposed method

• ACI uses a coarse-grained representation of ancestral relations (X → Y if X causes Y indirectly) to drastically reduce the search space compared to direct relation enumeration.
• It formulates ancestral reasoning rules in Answer Set Programming (ASP), encoding d-separation and independence constraints using logical rules and symmetry axioms.
• The method incorporates weighted independence statements from observational and interventional data via a loss function that penalizes inconsistencies with input data.
• Confidence scores are estimated as approximations to marginal probabilities of ancestral relations, derived from the consistency of evidence across multiple independence tests.
• A minimization objective is used to find the most consistent ancestral structure by minimizing weighted violations of input independence and dependence statements.
• The framework supports integration of background knowledge and can be extended to combine multiple data types through weighted constraints.

Experimental results

Research questions

• RQ1 Can ancestral reasoning reduce the computational complexity of constraint-based causal discovery while preserving or improving accuracy?
• RQ2 How can confidence scores be reliably estimated for causal predictions in the presence of noisy independence tests?
• RQ3 Can ACI effectively combine observational and interventional data to improve causal inference robustness?
• RQ4 To what extent does ACI outperform existing state-of-the-art methods in accuracy and speed on synthetic and real-world data?
• RQ5 How well does ACI handle faithfulness violations in real-world biological data, such as protein signaling networks?

Key findings

• ACI achieves a speedup of several orders of magnitude over the state-of-the-art method antti on synthetic datasets, significantly improving scalability.
• ACI outperforms standard methods like bootstrapped FCI and CFCI in terms of accuracy on synthetic data, particularly in the presence of latent confounders.
• The method successfully recovers causal structures from the challenging SPP05 protein signaling dataset, which previously resisted constraint-based methods due to faithfulness violations.
• Confidence scoring enables ranking of predictions, improving overall accuracy by prioritizing high-reliability relations.
• The integration of weighted ancestral constraints from combined observational and interventional data substantially increases prediction reliability.
• ACI is asymptotically consistent and sound under mild conditions on statistical tests, ensuring theoretical reliability.