acceptodds
Withdrawn from ICLR 2027

PRCD-MAP: Learning How Much to Trust Imperfect Priors in Causal Discovery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

External causal knowledge can improve structure learning, but incorrect or unevenly reliable priors can distort the learned graph. We propose PRCD-MAP, a framework that learns bounded trust parameters to modulate sparsity and ridge penalties in structural vector autoregression. Grouped temperatures calibrate trust across prior-probability bins; an optional contextual MLP assigns edge-specific trust. Trust updates use an empirical-Bayes-inspired agreement-and-curvature surrogate. Our analysis gives conditional estimation and local perturbation bounds, separating regularization robustness from causal identification and optimization guarantees. Experiments cover synthetic temporal graphs, CausalTime, electricity consumption, nonlinear and cross-sectional models, prior corruption, scalability, and component ablations. Cached data-driven priors improve AUROC over the no-prior backbone by 0.067 and 0.089 on AQI and Medical. In a controlled structured-error suite, learned trust reaches AUROC 0.850 versus 0.759 for a no-prior ablation, and distinguishes correct from incorrect prior assertions with AUROC 0.796. Fixed trust reaches 0.863 in that suite, and grouped temperatures lead a matched architecture comparison. Adaptive trust improves recovery in some tested regimes, but its benefits depend on sample size, prior reliability, and the optimization procedure.

Primary area:
causal reasoning
Keywords:
causal discovery, prior knowledge, trust calibration, structural vector autoregression, empirical Bayes

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_6QwXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes PRCD-MAP, a penalized structural VAR (SVAR) structure learner that uses an external edge-probability matrix as a soft prior. The prior modulates an edge-wise ℓ1\ell_1 weight on instantaneous edges and a prior-weighted ridge penalty. Its influence is controlled by temperature (trust) parameters on the prior's logits, either one per quantile bin of the prior (grouped trust) or one per edge from a small MLP (contextual trust). The temperatures are learned with an empirical-Bayes-inspired surrogate based on agreement between the calibrated prior and the current coefficient magnitudes. Experiments cover synthetic ER graphs with corrupted priors, the CausalTime benchmarks, a BA-graph study, and ablations.

Critical strengths
  1. The paper addresses a practically relevant problem with a simple mechanism.

Domain knowledge is often partial and unevenly reliable, and the proposed method offers a lightweight way to use such knowledge inside a standard SVAR learner. The ablation also suggests that using priors softly is considerably more robust than using them as hard constraints (Table 5).

  1. The learned trust tends to distinguish reliable from unreliable prior assertions.

The learned temperatures tend to increase with prior accuracy (App. M.1) and separate correct from incorrect prior assertions reasonably well (AUROC 0.796, Sec. 4.6).

Critical weaknesses
  1. Limited theoretical support and unclear novelty relative to prior-aware causal discovery.

Proposition 1 bounds a local perturbation from a neutral-prior solution under fixed support, order, and coefficient signs. Theorem 5 requires additional curvature and estimation conditions. Neither result establishes that the learned trust update attenuates an inaccurate prior or improves causal graph recovery. The paper’s central contribution is learning how much to trust a prior, but its theoretical results do not address the performance of that learning step.

Th Soft prior modulation and data-dependent reliability estimation have also been studied previously. The paper omits Hwang et al. (2026), Chen et al. (2025), and Amirkhani et al. (2017), and discusses Constantinou et al. (2023) only in connection with hard constraints. It does not establish what new principle the trust update contributes beyond this work or compare against an applicable prior-aware method using the same prior.

  1. Empirical evidence does not establish the benefit of learned trust with external priors.

The experiments are extensive, but they do not directly test the setting that motivates the paper. The controlled priors are generated by corrupting reference graphs, while the CausalTime priors are computed from correlations in the evaluated data. Thus, the results do not establish whether the method can learn the reliability of independently obtained external knowledge. Improvements over prior-agnostic baselines cannot by themselves isolate the benefit of learning trust.

The comparisons within PRCD-MAP also leave that benefit uncertain. Fixed trust outperforms learned trust at prior accuracy 0.9 for every sample size in Table 1 and in the structured-error suite. Low trust outperforms learned trust at prior accuracy 0.4 for every sample size in Table 29. These findings do not imply that learned trust is never useful, but the paper has not established when it offers an advantage over simpler trust choices.

Finally, Tables 1 and 15 report different results under the same stated setting, and several results used in the main text are marked as provisional because their per-run records were not recovered. This makes the empirical support for the central claim harder to assess.

Questions for authors
  1. For any tested identifiable SVAR setting, do the assumptions of Proposition 1 or Theorem 5 hold for the computed solutions? If so, what do they guarantee about graph recovery with learned trust?

  2. Why does agreement with coefficients fitted using the same prior reveal that prior’s reliability rather than reinforce its errors? What is the new principle beyond applying soft priors to an SVAR?

  3. With a prior obtained independently of the data and reference graph, how does learned trust compare with fixed and low trust under otherwise identical settings? Please report results by prior quality.

  4. What explains the different results in Tables 1 and 15 under the same stated setting? Can the provisional experiments be rerun with per-run records saved?

Reviewer_6ykuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper tries to solve how causal discovery use prior information when part of the prior distribution/knowledge is not accurate. The proposed model estimates a causal graph using data under some regularizations and update the graph based on the estimations. The authors evaluate the results on simulation studies based on some structural VAR data, CausalTime dataset, electricity data, and other nonlinear simulations. They compare the proposed model with sota causal discovery methods such as DYNOTEARS.

Critical strengths

When the reliability of prior knowledge is unknown, estimating the temporal relationship using graph and update the graph based on causal discovery is an interesting and promising idea.

The author has provided a lot of empirical studies on both real datasets and synthesized datasets to address problems such as prior corruption, sample size, and nonlinear relationships, and all these problems are practical and relevant to the main challenge of the partial reliability of prior knowledge. Meanwhile, ablations show the component benefit such as add or remove prior information, remove trust update, remove cuvature term and etc.

The theories show the effectiveness and limitation of prior regularization includng the sensitivities to prior, optimization problems and the corresponding assumptions.

Critical weaknesses

The theoretical contribution is a bit incremental

The theoretical contribution is relatively incremental. In prop 1, the paper shows ∥wτ−w0∥2≤bound\|w_\tau-w_0\|_2\leq bound This is like a standard strong-convexity argument to show how change regularization will effect the local estimation, but it does not address the central problem wheter the proposed trust update learns to weight the incorrect part of the prior knowledge.

The main theory section mainly says a story that if trust is small and the local assumptions hold, then the estimate will stay close to the prior estimate. It does not show how the proposed update actually learns the small trust for incorrect priors and this is the main challenge of the paper.

The method combines prior knowledge based on weights regulizatons, tem scaling, optmization and a MLP, but the necessity of each independent component is not clearly explained. In other words, why or what the proposed specific trust objective offers the benefit beyond simple adaptive regulization?

The assumptions are stated clearly, but the paper does not show that the estimated graphs with and without prior influence share the same support, ordering, and coefficient signs, as required by the local analysis.

Questions for authors
  1. How does the proposed trust update learn to downweight incorrect priors, beyond the local sensitivity guarantee in Proposition 1?
  2. What benefits do the curvature term and contextual MLP provide over simpler grouped trust, given the mixed ablation results?
Reviewer_p3oDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper learns causal graphs from time series data with a weighted prior SVAR. The main contribution is learning how much to trust each prior via pair-bin temperatures as well as a heuristic to reward agreements between prior and current graphical estimation. In synthetic and time series data, the informative priors seem to help, but barely beat simply fixing the trust or dropping the prior.

Critical strengths

The problem of learning how to trust priors and fixing prior strengths is a very important problem in the field of causal learning.

Critical weaknesses

The paper does not look like a normal paper. Instead, it is 48 pages of mostly trivial mathematical lemmas, including many appendices (A through Z and beyond). In my opinion, this does not qualify as a paper ready to be published, nor does it include any informative and important contribution. In my opinion, the contribution is minimal and trivial, despite all the formulations.

Questions for authors

How were the CausalTime series of length T=400 constructed? The benchmark releases 500 samples of 40 steps and its reference graph is lagged-only. Please also explain why DYNOTEARS differs between Table 2 (.503/.504/.512, zero variance) and Tables 6 and 35–37 (.554/.583/.521) on the same data, and why learned-τ at acc 0.4, T=500 is 0.828 in Table 1 but 0.893 in Table 29.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.