acceptodds
Withdrawn from ICLR 2027

Learning Causal Variable Orderings with Pointer Networks for Causal Structure Discovery

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Causal structure learning from observational data requires searching over a large combinatorial space of possible directed acyclic graphs (DAGs), making accurate and efficient structure learning challenging. Existing causal discovery methods based on continuous optimization and reinforcement learning either optimize graph structures directly or explore variable orderings through sequential decision making. We propose CaPo, a neural causal discovery framework that formulates causal variable ordering as a learnable permutation to guide constrained graph search. A Pointer Network generates candidate variable orderings, while an ordering-induced mask restricts admissible edge directions according to the learned ordering, thereby reducing the search space over candidate graph structures. The key novelty is to learn variable orderings directly from observational data and optimize them according to downstream causal graph quality, rather than relying on fixed or explicitly searched orderings. The ordering policy is optimized through reinforcement learning using a graph-level causal structure reward, and a DAG-constrained graph learner estimates the corresponding causal structure under the learned ordering. Experiments on Bayesian network benchmarks of varying sizes show that CaPo achieves competitive or lower Structural Hamming Distance (SHD) and higher F1-scores than state-of-the-art causal discovery baselines across multiple datasets. CaPo also requires less computational time than several compared methods, demonstrating efficient causal structure learning across networks of different sizes.

Primary area:
causal reasoning
Keywords:
Causal Structure Learning, Causal Ordering, Pointer Networks, Reinforcement Learning, Directed Acyclic Graphs

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

CaPo uses a Pointer Network to generate causal variable orderings. An ordering-induced mask and a screening mask constrain a continuous graph learner, whose graph-level reward updates the ordering policy through REINFORCE. The paper reports results on Bayesian network benchmarks with up to 724 variables.

Critical strengths

1.Coupling ordering search with continuous graph estimation is a reasonable approach to reducing the search space.

2.The evaluation covers a broad range of graph sizes and reports ordering, reward, and runtime comparisons.

Critical weaknesses

1.The main metrics appear inconsistent. For the standard 592-edge PIGS network, the reported recall of 0.5533 implies about 264 missing true edges, already exceeding the reported SHD of 168. Different graph preprocessing or metric definitions could explain this, but neither is specified.

The bnlearn repository lists 592 arcs for the 441-node PIGS network. If Table 1 uses this graph and standard directed-edge recall, recall 0.5533 implies approximately 592 × (1 − 0.5533) = 264.45 missed true edges on average. Standard directed-graph SHD cannot be smaller than that count, even when a reversed edge costs one edit. The reported SHD is 168. A different reference graph, graph preprocessing, or metric convention could invalidate this comparison. The concern is therefore a concrete request to reconcile definitions and results, not a finding of fabrication or evidence about AI authorship.

2.The distinction from CORL is insufficiently established. CORL already uses a pointer-based ordering policy, reinforcement learning, and an episodic graph-score reward. The main possible advance is the coupling to the continuous graph learner, which needs clearer isolation. See Wang et al., IJCAI 2021, Sections 4.1–4.2.

3.The role of the acyclicity penalty is unclear. A strict ordering-induced mask already makes the masked graph acyclic. Figure 1 nevertheless computes a DAG penalty from that graph. The paper needs to specify which matrix is penalized and explain how this term affects learning.

Questions for authors

1.Can the authors provide the exact PIGS reference graph, metric implementation, per-seed predictions, and the numerical traces and plotting procedure behind Figure 4?

2.Which matrix enters each graph-loss term, and what does the continuous graph learner add beyond an otherwise matched CORL-style ordering policy?

Reviewer_26vhOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes CaPo, a causal discovery framework using Pointer Networks and reinforcement learning. At each reinforcement learning step, CaPo first generates possible variable permutations by Pointer Network as causal edge constraints. Then the graph learner estimates the causal structure with DAGMA. Finally, CaPo calculate reinforcement reward according to learned causal structure and update the ordering policy. Experiments on Bayesian Network benchmarks from the Bayesian Network Repository are provided to the effectiveness of CaPo. However, the paper lacks detailed implementation of the proposed Method.

Critical strengths

The paper integrates reinforcement learning into causal discovery with permutation modeling, which is a meaningful direction. It is worth investigating the further potential usage of reinforcement learning in causal area.

Critical weaknesses

1.The paper provides almost no implementation details of CaPo, with almost all module introduction at a high level. The details of Pointer Network architecture, REINFORCE formulas and other modules are never given in the paper, with the Method section only one page length. Appendix also contains no implementation of CaPo. 2. The novelty of CaPo is questionable. Causal discovery with reinforcement learning using variable permutation has already been studied, and using neural network for permutation module is also common. For example, in the mentioned methods CORL and RCL-OG. Therefore, the claimed contributions of CaPo in fact lack innovations.

Questions for authors
  1. Could the authors provide detailed implementation of CaPo, so the method can be properly understood and reproduced?
  2. Could the authors further clarify the novelty of the paper especially compared to those reinforcement learning causal discovery methods?
  3. The method figure shows several errors. Is it generated by AI and without properly revised for correctness?
  4. The experiments are using graphs from Bayesian Network Repository. Could the authors provide more experiments on synthesis data with systematically randomly generated graphs, and provide real-world data experiments?
Reviewer_VzVjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CaPo, a causal discovery framework that uses a Pointer Network to learn variable orderings through REINFORCE and constrains DAGMA-based graph learning using the generated permutations. Empirical studies on Bayesian network benchmarks demonstrate improved Structural Hamming Distance (SHD) and reduced runtime compared with several existing methods.

Critical strengths

The integration of neural ordering with constrained graph learning achieves competitive SHD and computational efficiency across different graph sizes.

Critical weaknesses

W1. The paper does not explain under what assumptions the graph-quality reward can recover the true causal ordering from observational data, particularly when multiple Markov-equivalent DAGs exist.

W2. Despite improved SHD, recall drops to 0.0482 on Pathfinder and 0.1049 on Link, raising concerns about the claimed scalability and practical effectiveness of causal structure recovery.

W3. The correlation-based preprocessing can eliminate genuine causal edges with weak marginal correlations, but its effect on graph recovery is not adequately investigated.

Questions for authors

Q1. What assumptions ensure that optimizing the graph-quality reward identifies the correct causal ordering rather than an observationally equivalent alternative?

Q2. Since the ordering mask already enforces acyclicity, what additional benefit does the DAGMA acyclicity penalty provide?

Q3. How does performance change when correlation-based screening is removed, particularly on larger or nonlinear causal graphs?

Q4. Can the authors explain the substantial decline in recall and F1-score as graph size increases despite improvements in SHD?

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

Related papers

Discussion (0)

Sign in to comment.