Interpretable Causal Discovery via Causal-Effect Constraints

Cixuan Zhang, Guy Van den Broeck, Benjie WangarXiv 2026 · arXiv

TL;DR

This paper introduces a novel framework for conditional causal discovery, allowing researchers to identify causal structures that explain specific, user-defined constraints (such as an unusually large causal effect between two variables). By adapting rare-event simulation techniques, the authors overcome the computational failure of standard Bayesian methods when dealing with events that have tiny posterior probabilities, enabling interpretable, pathway-level "what-if" scientific analysis.

What Problem Does This Paper Address?

Causal discovery algorithms typically try to answer a broad question: What is the most probable causal graph given our observational data? However, in scientific exploration, researchers often want to perform targeted "what-if" analyses. For instance, a biologist looking at protein signaling might ask: What specific biological pathways could explain an unusually strong downstream effect from Protein A to Protein B?

Answering this requires conditional causal discovery—estimating a Bayesian posterior over graphs and parameters conditioned on a specific causal-effect constraint. The core challenge is computational: these specific constraints often represent "rare events" with extremely small posterior mass under the unconstrained distribution. Standard Markov Chain Monte Carlo (MCMC) samplers rarely, if ever, visit these regions, and simple rejection sampling becomes hopelessly inefficient (the "zero-hit" problem), leaving researchers unable to explore alternative explanatory pathways.

Methods

The authors address this challenge by framing conditional causal discovery as a rare-event posterior inference problem. Their approach consists of two main components:

  1. Score-Level Formulation: The user's scientific query is converted into a scalar score function $h(Z)$, where $Z = (G, B)$ represents the joint state of the graph structure $G$ and edge weights $B$. A constraint is satisfied when $h(Z) \ge \lambda_\star$. This flexible formulation easily accommodates single-effect constraints, left-tail (negative) effects, or complex conjunctions of multiple constraints.
  2. Adaptive Multilevel Splitting (AMS): Instead of trying to sample the rare event directly, the algorithm iteratively drives a population of particles toward the target region. It automatically discovers a sequence of intermediate, increasingly strict thresholds. At each level, particles that fail to clear the threshold are discarded and replaced by resampled versions of the survivors, which are then diversified using an inner MCMC mutation kernel.

The inner MCMC kernel operates over the joint graph-parameter space, alternating between structure moves (using standard Structure-MCMC or the advanced PARNI-DAG sampler) and blocked parameter refreshes, ensuring the sampler doesn't get stuck in local structural minima.

Figure 1: Overview of conditional causal discovery. Starting from the unconstrained Bayesian posterior over causal graphs and edge weights, we condition on a user-specified extreme causal-effect event and obtain a constrained posterior whose samples can be summarized at the pathway level.

Datasets

The framework was evaluated using two main setups:

  • Synthetic Benchmarks: Linear-Gaussian datasets generated from random Erdős-Rényi DAGs across varying dimensions ($d \in \{4, 8, 16, 32\}$), each with 1,000 observational samples.
  • Sachs Protein-Signaling Dataset: A classic real-world benchmark in causal discovery consisting of $n = 7,466$ continuous measurements of $d = 11$ proteins, paired with a widely accepted 20-edge consensus network used as a qualitative reference.

Key Insights and Ideas

  • Logarithmic vs. Linear Scaling: The primary algorithmic "aha moment" is that by breaking a single rare event into a sequence of nested conditional steps, the computational cost scales logarithmically with the rarity of the event ($1/p_\star$), rather than linearly. This makes it possible to accurately estimate posterior probabilities as small as $10^{-7}$ or lower.
  • Constraints as a Regularizer: When multi-effect constraints are constructed around ground-truth values, conditioning on these constraints actually causes the posterior to concentrate tightly around the true data-generating mechanism, reducing structural uncertainty (Structural Hamming Distance).
  • Pathway-Level Interpretability: In under-determined settings like the Sachs dataset, the unconditioned posterior often spreads its mass across many conflicting or zero-effect graphs. Forcing an extreme-effect constraint cleanly isolates the dominant direct and mediated pathways that can physically support that effect.

Key Results

  • Ground-Truth Validation ($d=4$): The method's tail-probability estimates were validated against an exact, exhaustive enumeration of all 543 possible 4-node DAGs. The AMS framework closely matched the exact reference baseline across 10 distinct test cases, whereas a prominent baseline (DiBS) overestimated tail probabilities by multiple orders of magnitude.
Figure 2: Single-effect one-sided tail-probability estima- tion on d = 4 across 10 test cases, with exhaustive enumer- ation as the reference posterior probability. Error bars show across-run variability over 10 independent runs. DiBS is omitted from the plot because its estimates are orders of magnitude larger on these tail events, which would obscure the comparison among the remaining methods; the corre- sponding numerical values are reported in Table 3.
  • Scalability up to $d=32$: On larger synthetic graphs where unconditional baselines completely failed to generate valid samples, the proposed method produced smooth, stable, and reproducible monotone-decaying tail curves. Integrating the locally informed PARNI-DAG structure kernel provided significantly better stability in the deepest tails compared to basic Structure-MCMC.
  • Sachs Case Study: Under a joint extreme-effect constraint on two separate protein pairs (PIP3 $\rightarrow$ PIP2 and Erk $\rightarrow$ Akt), the sampler successfully uncovered a shared biological mediator (Plcg) that was entirely obscured under standard, unconditioned posterior sampling.

Limitations and Future Work

The primary limitation of the current implementation is its reliance on linear-Gaussian Structural Equation Models (SEMs) and the assumption of causal sufficiency (no unobserved latent confounders).

To make this framework applicable to a broader range of real-world scientific problems, the authors note that future work must extend the inner architecture to support:

  1. Nonlinear mechanism classes (such as splines, Gaussian Processes, or neural networks).
  2. Latent confounding and the integration of interventional data.
  3. Richer constraint families, such as path-specific effects or qualitative monotonicity constraints.

Summary updated

Abstract

Causal discovery aims to uncover the underlying causal relationships given data generated from a system. The goal, however, is not merely to predict causal edges given data, but also to be able to interpret and explain either observed or hypothesized phenomena, such as a particularly large causal effect. We consider this task of conditional causal discovery and cast it as a Bayesian inference problem, in which we target the posterior over causal graphs and parameters conditional on an event such as a causal-effect constraint. Unfortunately, this poses a computational challenge: existing approaches to Bayesian causal discovery struggle when the event has small posterior mass. To address this, we adapt rare-event estimation techniques to perform inference the joint graph-parameter space. Our method gradually drives a particle population toward the constrained region while maintaining samples that approximate the conditional posterior. Empirical evaluation on synthetic graphs validates the accuracy of our approach at small and large scales, and we show in a case study on the Sachs protein dataset how our method can be used to aid scientific exploration by providing pathway-level summaries.
BibTeX
@misc{zhang2026interpretable,
  author = {Cixuan Zhang and Guy Van den Broeck and Benjie Wang},
  title = {{Interpretable Causal Discovery via Causal-Effect Constraints}},
  year = {2026},
  eprint = {2608.12640},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2608.12640},
}
Figure