acceptodds
Under review as a conference paper at ICLR 2027

Mesoscopic Causal Discovery: Learning Target-Centric Causal Subgraphs via Differentiable Optimization

Abstract

Causal discovery methods have traditionally focused on either recovering global causal structures or identifying local relationships around a target variable. However, many real-world applications require understanding causal mechanisms at an intermediate scale, where the goal is to uncover a structured subset of variables that are causally relevant to a target through both direct and multi-step pathways, without incurring the cost of full graph reconstruction. In this paper, we study Mesoscopic Causal Discovery in the form of target-centric k-hop causal subgraph recovery, bridging global and local causal learning. We formalize this setting as Target-Centric Causal Subgraph Discovery, which aims to recover a directed acyclic target-centric subgraph within a specified k-hop structural scope, capturing both immediate dependencies and the intermediate pathways through which causal influence propagates. This problem is challenging because extending beyond immediate neighborhoods rapidly enlarges the candidate space, while indirect causal pathways can be difficult to distinguish from spurious associations. To address these challenges, we propose Target-Conditioned Causal Subgraph Learning (TCSL), a differentiable framework that constructs a target-dependent candidate space from observational data and adaptively refines variable relevance through stochastic node gating while jointly learning causal structure and conditional distributions. TCSL combines conditional-flow modeling with a differentiable acyclicity constraint to ensure valid causal structures and extracts the target-centric k-hop causal subgraph from the learned structure without access to ground-truth neighborhood information. Extensive experiments on both synthetic and real-world datasets demonstrate the effectiveness of TCSL.

Primary area:
causal reasoning
Keywords:
causal discovery, causal structure learning, target-centric causal subgraphs, differentiable optimization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_tQAuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies target-centric kk-hop causal subgraph recovery and proposes TCSL, combining correlation-based screening, stochastic node gating, higher-order structural mixing, conditional affine flows, and an acyclicity constraint. The method achieves a favorable accuracy–cost trade-off in some settings. However, gaps remain between the problem motivation, the causal interpretation of the method, and the experimental evidence.

Critical strengths
  1. The problem has practical relevance. Recovering intermediate-scale structures around a target could combine the efficiency of local discovery with the structural coverage of global methods.
  2. The task and inference pipeline are clearly specified. The inputs, query radius, and output subgraph are defined explicitly. Neighborhood extraction operates on the predicted graph without requiring ground-truth neighborhood information.
  3. Some empirical results are promising. In several synthetic settings, TCSL reports better structural recovery than local baselines, while its runtime is lower than the full-graph training cost of the compared global methods.
Critical weaknesses
  1. The problem definition is not fully aligned with the motivation. The introduction emphasizes multi-step causal influence, whereas the task defines neighborhoods using undirected skeleton distance. Such neighborhoods can contain nodes with no directed causal path to or from the target. The authors should clarify whether the objective is structural neighborhood recovery or causal mechanism recovery, and connect each motivating challenge to a corresponding methodological design.

  2. The substantive contribution of the combination remains unclear. The paper does not establish why this particular combination is necessary or what property it provides beyond a straightforward integration of existing components. Controlled comparisons, particularly existing structure learners operating on the same candidate sets, are missing. Consequently, the contributions of screening, gating, and higher-order mixing remain difficult to separate.

  3. The method’s correspondence to direct causal edges is insufficiently justified. Higher-order mixing in Eqs. (10)–(12) allows non-parent ancestors to enter a node’s conditional model, yet A is interpreted as the adjacency matrix of direct causal relations. Acyclicity and valid density evaluation alone do not establish this interpretation. Moreover, screening may exclude common causes, so causal sufficiency of the full variable set need not hold within the candidate set.

  4. The evidence does not adequately cover the multi-hop claims or their boundaries. The main ablations and real-world evaluation are restricted to k=1. Reported error bars capture variation across targets rather than independent graphs or random seeds, and edge-level metrics do not directly establish pathway recovery.

Questions for authors
  1. What is the precise recovery objective? Please distinguish structural neighborhoods from directed causal pathways and explain how the chosen definition addresses the mechanism-analysis needs stated in the introduction.

  2. What unique property does the combination provide? Which component distinguishes indirect causality from spurious association, and why would screening followed by an existing method be insufficient?

  3. How is the causal interpretation of the output justified? When higher-order mixing introduces non-parent ancestor information and screening may omit common causes, why can A still represent direct causal structure?

  4. Can the experiments directly test the central claims? Please consider comparisons using identical candidate sets, ablations at k=2,3, and separate metrics for candidate coverage, node-set recovery, and directed-path recovery.

Reviewer_mhE2OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the recovery of target-centric kk-hop causal subgraphs from observational data and proposes TCSL for this task. The method combines Spearman correlation-based screening, stochastic node gating, conditional distribution modeling, and differentiable DAG constraints to learn causal structures. Experiments show competitive results across several synthetic settings, but its ability to recover multi-hop causal subgraphs from real-world data and its robustness remain insufficiently validated.

Critical strengths
  1. Meaningful problem formulation. The focus on multi-hop causal structures around a target variable offers a useful direction between local causal discovery and full-graph structure learning.
  2. Well-integrated methodology. TCSL combines candidate screening, stochastic node gating, conditional distribution modeling, and DAG constraints within a unified optimization framework.
  3. Broad synthetic evaluation. Experiments cover two nonlinear data-generating mechanisms, graphs with up to 500 nodes, different sample sizes, and multiple hop budgets, with an additional evaluation on the real-world Sachs dataset.
Critical weaknesses
  1. Limitations of candidate screening. Spearman correlation screening, the top-bb selection strategy, and the candidate-set size cap may exclude causally relevant variables that subsequent optimization cannot recover. The paper does not report candidate-set coverage of the ground-truth kk-hop neighborhood or provide a sensitivity analysis of the screening parameters.
  2. Limited expressive capacity of the conditional model. Equation (11) aggregates information from multiple variables into a single scalar before feeding it into a node-wise MLP. This design may struggle to capture complex nonlinear interactions among multiple parents, potentially affecting causal structure recovery. The paper provides no dedicated experiments to evaluate this limitation.
  3. Lack of robustness evaluation under unobserved confounding. The method relies on assumptions such as causal sufficiency and independent exogenous noise, but the paper does not evaluate its performance in the presence of latent confounders or omitted variables. Its reliability when these assumptions are violated therefore remains unclear.
  4. Insufficient validation of multi-hop discovery on real-world data. The Sachs dataset experiments consider only k=1k=1. Consequently, the paper's central claim regarding multi-hop causal subgraph recovery is supported primarily by synthetic experiments, with no direct validation on real-world data.
  5. Incomplete ablation studies of key components. The existing ablations mainly examine stochastic node gating and structural mixing order. The individual contributions of candidate screening and scalar aggregation are not separately evaluated, making it difficult to determine how much each component contributes to the overall performance.
  6. Insufficient theoretical guarantees. The paper assumes that the data-generating mechanisms fall within an identifiable setting, but does not clearly specify the corresponding identifiability conditions. It also provides no consistency analysis or theoretical results establishing under what conditions the proposed optimization objective can recover the true causal structure.
Questions for authors
  1. How does the scalar aggregation in Equation (11) handle nonlinear interactions among multiple parents? Could the authors provide additional experiments involving causal mechanisms with explicit nonlinear interactions between multiple parent variables?
  2. Could the authors repeat the experiments across multiple independently generated DAGs and randomly selected target variables, reporting performance variability across graph instances and random seeds?
  3. How are the local baselines adapted to the multi-hop subgraph recovery task when k>1k>1? Could the authors also report runtime measurements under a consistent evaluation protocol, along with the necessary hardware specifications and implementation details?
Reviewer_6CbeOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes TCSL, a target-centric causal subgraph learning framework for mesoscopic causal discovery. TCSL method first constructs a target-dependent candidate set and then uses affine flow to model data, utilizing node-specific MLP to predict detailed location and scale affine flow parameters. The model and causal relationships are learned by aligning the predicted conditional Gaussian distribution and observed data, using negative log-likelihood by ALM optimization. Finally, TCSL extracts a mesoscopic causal graph by thresholding the graph and then extracting the subgraph within a certain number of hops from the target. Experiments on synthetic datasets and Sachs are conducted to verify the effectiveness of proposed method.

Critical strengths
  1. Mesoscopic causal discovery is a meaningful research direction, which completes current vacancy between global and local causal discovery methods.
  2. The paper has designed a relatively complete experimental plan including synthetic and real dataset, with comparative and ablation experiment settings.
Critical weaknesses
  1. Although TCSL claims to solve mesoscopic causal discovery, the target-centric subgraph is only extracted at a post-processing stage. The model is still learning the causal graph over a large variable set without mesoscopic constraint. The methodological novelty of TSCL in mesoscopic causal discovery is unclear compare to applying same preprocessing and extracting strategy to existing causal discovery methods. 2.The paper lacks correctness analysis connecting distribution fitting to causal discovery using conditional Gaussian likelihood optimization. What’s more, the article also lacks real analysis of mesoscopic causal discovery identifiability but only assume an identifiable global structure.
Questions for authors
  1. Could the authors clarify how current design is specifically suitable for mesoscopic causal discovery?
  2. Could the authors provide stronger correctness analysis of TSCL and identifiability of mesoscopic causal structure learning?
  3. TSCL is evaluated by Precision, Recall and a custom Distance metric. Could the authors provide more standard metrics such as F1, AUPRC and SHD?
  4. TSCL use Gaussian distribution to model affine flow. Could the author provide more theoretical and experimental analysis of data generated by non-Gaussian noise distributions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.