acceptodds
Under review as a conference paper at ICLR 2027

Integrating Overlapping Static and Timeseries Causal Graphs via Constraint-Solving

Abstract

In many causal learning problems, variables of interest are not always measured in the same observations, but are instead distributed across multiple datasets with overlapping variables. Previous algorithms used independence tests to directly determine the minimal equivalence class of DAGs consistent with all input graphs, but these methods faced significant computational limits. In this paper, we formulate this problem as a more computationally efficient answer set programming (ASP) problem called SCION (Solving Constraints for Integration of Overlapping Networks), which can be solved with the ASP system *clingo*. We then extend SCION to a novel method, SCION-TS, for solving the overlapping variables problem in the timeseries setting. We test both algorithms across multiple simulations to demonstrate their computational viability, and successfully apply both algorithms to input data, using p-value pooling to improve input graph estimation.

Primary area:
causal reasoning
Keywords:
Causal learning, constraint satisfaction, answer set programming, timeseries

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

The paper proposes SCION, an ASP formulation of the overlapping causal graph integration problem addressed by ION, using the underlying optimizations of the clingo solver for more efficient computation. It further extends the formulation to the time-series setting with Markov order 1 ground-truth processes. Experiments demonstrate improved computational efficiency over ION and the benefits of incorporating time-lag information.

Critical strengths

The simulations demonstrate the computational feasibility of using clingo to scale SCION to sparse graphs of moderate size, and suggest that SCION-TS can also handle time-series instances. This has potential practical value.

Critical weaknesses

The theoretical rigor of the paper is seriously limited. Examples include:

  • The treatment of bidirected edges and edges with circle endpoints in Listing 1.

    • Line 20 of Listing 1 is only a rule for deriving bidirected. When bidirected is supplied as an input fact, this rule does not require the candidate DAG's MAG to realize that bidirected edge, including the inducing path and the non-ancestral relationships at both endpoints in the candidate DAG

    • Listing 1 provides no constraints requiring input facts expressed using causalconn and undirected_causalconn to match relationships induced by the candidate graph. A candidate can therefore satisfy the program without the required connection.

      • For example, consider a single input PAG X∘−∘YX\circ - \circ Y, encoded as an undirected_causalconn fact according to the description in lines 143–146. Listing 1 admits the empty candidate DAG, in which X⊥YX \perp Y. This contradicts the d-connection preservation claimed in Theorem 1.
    • For X∘→YX\circ \to Y, the candidate DAG must realize the corresponding adjacency in the MAG and satisfy Y∉An⁡(X)Y\notin \operatorname{An}(X). Circle endpoints also cannot be completed independently edge by edge. The resulting MAG must belong to the Markov equivalence class represented by the input PAG. These conditions are neither analyzed nor implicit in the paper's current treatment.

  • The paper conflates DAG, MAG, PAG, and latent-projection ADMG.

    • The output graphs Hi\mathbb{H}_i are defined as DAGs in lines 104–105, whereas Theorem 2 quantifies over a PAG Hi\mathbb{H}_i and concludes that it belongs to the output set.

    • Lines 196–198 directly call the result of marginalizing a known DAG a PAG, using X→Z→YX \to Z \to Y becoming X→YX \to Y as an example. First, if the graph in this example is intended to be a PAG, the orientation between XX and YY should be undetermined. Second, this operation could yield either a latent-projection ADMG or a MAG, and the authors’ current description is unclear.

  • The proofs in the appendix lack rigor and clarity. Examples include:

    • Lines 646–649: Line 22, together with the recursive definitions, excludes certain directed paths and common-cause paths whose internal nodes are unobserved in the relevant input. This does not directly exclude general active paths under arbitrary conditioning sets. Moreover, activating a collider by conditioning on its descendant does not imply that directed paths exist from both endpoints to that descendant. The appeal to the recursive directed rules therefore does not establish the claimed contradiction.

    • Lines 651–656: Line 23 only constrains definite directed input edges. The proof does not explain how it covers bidirected edges, edges with circle endpoints, or general connection relations under arbitrary conditioning sets.

    • Lines 664–665: The fact that clingo enumerates all graphs satisfying the encoded constraints does not by itself establish completeness for the original problem.

[1] Ancestral graph Markov models. Ann. Statist. Thomas Richardson, Peter Spirtes. 2002.

Questions for authors

Could the authors provide a more rigorous and clearer graph-theoretic discussion (for example addressing the issues raised in the Weaknesses section)?

Reviewer_jwr1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes to cast the problem of causal structure learning from multiple datasets that have overlapping variables as an answer set programming problem. The paper develops an ASP formulation known as SCION and uses an existing ASP solver named clingo to solve it. It then extends its method for time-series data. The paper gives a sound and complete algorithm. The paper provides some synthetic experiment results.

Critical strengths
  • The paper provides an ASP formulation for solving a causal structure learning problem motivated by practical scenarios.
  • The paper proves the soundness and completeness for both static and time-series case for the problem of integrating overlapping networks under ASP formulation.
  • I find that being able to prove that Listing 2 is a direct encoding of the SCION-TS quite novel. The paper provides a good discussion (Lines 315-322) why the Markov order 1 assumption is needed.
  • The experiment substantiates the efficiency of the proposed ASP scheme by showing results on 25 node DAGs along with imperfect PAG inputs based on synthetic data. It also compares ION with SCION as an appropriate baseline and report the results in Appendix B. The experiment also examines SCION with real-world data based on European Social Survey
Critical weaknesses

Clarity and correctness:. The wording of the theorems on page 3 seems misleading. The soundness of the algorithm should be: If XX and YY are d-separated given ZZ in a Hi∈HH_{i} \in \mathcal{H}, there exists a Gi∈GG_{i}\in \mathcal{G} such that X and Y are d-separated given Z in GiG_{i}. For completeness, it should be: If XX and YY are d-separated given ZZ in a Gi∈GG_{i} \in \mathcal{G}, XX and YY are d-separated given ZZ in HiH_{i} for all Hi∈HH_{i} \in \mathcal{H}. The current proof of soundness is basically the proof of completeness. For lines 649-650, it is not clear how the paths from X to R or Y to R necessarily contradict the known d-separation in $\mathcal{G}_{i} $ .

However, the only d-separation stated in GiG_{i} is XX is d-separated with YY given ZZ in the proof. The proof should elaborate more on this. Besides, I also find it very difficult to follow the proof because it requires the reader to understand the logic of Listing 1 based on the clingo syntax at each line beyond the provided descriptions in Lines 139-149. Also, more discussion should be provided to justify the assumption used for time-series data. This is the main reason I currently think the paper sits at the borderline. Once that is addressed, I have no issues with accepting this paper.

Questions for authors

Clarification:

  • Is SCION-TS more efficient than SCION since more constraints can be introduced with directed arrows in general? Are there any difficulties in applying SCION directly on time-series data besides encoding it under Markov order 1 assumption?
  • Can the authors address the critical weakness I put about the clarity and correctness?

Suggestion for future work:

  • Have the authors thought about how to approach this problem if some side information are used to derive some input PAGs?
Reviewer_zu76OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to recover all possible global causal structures from multiple local causal graphs with partially overlapping variables. The authors propose SCION, which formulates the integration of static causal graphs using Answer Set Programming and leverages the clingo solver to improve computational efficiency. Building on this framework, they further introduce SCION-TS, extending the approach to time-series settings.

Critical strengths
  1. Partially overlapping variable sets across multiple datasets are common in real-world scenarios, making this problem practically important.
  2. By leveraging an ASP solver, SCION significantly reduces computational overhead compared with ION in the experiments.
Critical weaknesses
  1. Prior work has already employed SAT-based constraint solving to integrate local causal graphs. The core idea of this paper is similar to these existing approaches. The authors should further clarify the substantive contributions of the ASP formulation and the time-series extension.
  2. The proposed approach heavily relies on local causal graphs generated by underlying algorithms such as FCI and LPCMCI. Errors in local causal relationships may propagate to the global structure, potentially causing all candidate graphs to share the same incorrect relationships. Although the authors discuss this issue, a systematic analysis of error propagation mechanisms and their impact on global structure recovery performance is still lacking.
  3. In many experiments, SCION-TS requires removing input edges to obtain feasible solutions. However, this repair process may result in the loss of true causal information. The paper lacks sufficient validation of the effectiveness of the repair strategy, as well as systematic robustness experiments under varying levels of input errors, sample sizes, and underlying causal discovery algorithms.
  4. SCION-TS assumes that all direct causal edges in the true global causal graph have a one-step time lag. This assumption limits its applicability to settings involving higher-order time lags or contemporaneous causal relationships.
  5. The static method is primarily compared with ION, without direct comparisons against existing SAT-based causal graph integration methods. Similarly, SCION-TS lacks systematic comparisons with existing time-series causal integration approaches, making it difficult to fully assess its relative advantages.
  6. The paper computes accuracy by combining predictions of both edge presence and edge absence, which may obscure poor recovery of true causal edges. It would be more informative to report Precision, Recall, and F1 separately for causal edge recovery.
Questions for authors

Please refer to the weaknesses listed above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.