acceptodds
Under review as a conference paper at ICLR 2027

Robust Causal Discovery via Abductive Quantitative Argumentation

Abstract

Constraint-based causal discovery is notoriously fragile under finite-sample errors, which often trigger devastating structural error cascades. To address this, we propose Abductive Quantitative Argumentation for Causal Discovery (AQACD), a probabilistic-symbolic framework that reformulates causal learning as abductive argumentation. Unlike deterministic algorithms or exact logic solvers, AQACD models causal relations as defeasible hypotheses, dynamically evaluated through a continuous dialectic network of statistical and structural arguments. Rather than making irreversible commitments, graph revisions are guided by Inference to the Best Explanation: the topology is updated only when an alternative explanation decisively outweighs its rivals under a dialectic margin, rescuing mistakenly pruned edges (). This approach strictly improves global internal consistency while bypassing the exponential bottlenecks typical of exact logical inference. Extensive experiments on standard benchmarks demonstrate that AQACD consistently outperforms classical and continuous baselines. Most notably, it achieves substantial gains in orientation accuracy (Arrow F1) within Markov Equivalence Classes, exhibiting exceptional robustness to noise and neutralizing error cascades.

Primary area:
causal reasoning
Keywords:
Causal Discovery, Abductive Reasoning, Quantitative Argumentation, Conditional Independence Tests, Bayesian Networks

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_vw42OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Abductive Quantitative Argumentation for Causal Discovery (AQACD) to mitigate structural errors in causal graphs caused by erroneous conditional independence (CI) tests under finite-sample settings. The method integrates statistical evidence, graph-structural constraints, and orientation preferences, dynamically adding or deleting edges through non-monotonic graph revision. Experiments on five discrete Bayesian network benchmarks show improvements in skeleton recovery, orientation accuracy, and some intervention-related metrics.

Critical strengths
  1. The research problem is meaningful. Allowing the algorithm to recover erroneously deleted edges provides a promising approach to mitigating error propagation in conventional constraint-based causal discovery.
  2. The method is reasonably comprehensive. It combines quantitative argumentation, abductive reasoning, and global structural constraints, supported by ablation studies and theoretical analysis.
  3. The paper conducts repeated experiments, statistical significance tests, and component ablations across multiple benchmarks. It also evaluates orientation accuracy and intervention-related performance.
Critical weaknesses
  1. Potential inconsistency in the theoretical assumptions and default parameters. The asymptotic consistency theorem requires the regularization parameters to satisfy [ \lambda \max(w_{\mathrm{sparse}}, w_{\mathrm{coll}}) < \gamma, ] where the faithfulness separation margin (\gamma) is bounded by the range of the dependence evidence, implying (\gamma \leq 1). However, the default settings reported in the paper are (\lambda = 1) and (w_{\mathrm{coll}} = 1.5). These values imply [ \lambda \max(w_{\mathrm{sparse}}, w_{\mathrm{coll}}) \geq 1.5 > 1 \geq \gamma, ] which contradicts the stated condition. This raises concerns about whether the theoretical guarantee applies to the default configuration used in the experiments. The authors should clarify or correct this discrepancy. Moreover, the asymptotic analysis assumes a sample-size-dependent significance threshold, whereas the experiments use a fixed threshold, and the relationship between these settings should be explained.
  2. The independent contribution of the dynamic revision mechanism is not fully established. Ablation results suggest that its benefits vary across datasets and evaluation metrics. A more direct analysis of the trade-off between correctly recovered edges and incorrectly added edges is needed.
  3. The experimental scope is limited. The evaluation focuses primarily on five discrete Bayesian networks, and some baselines fail to run on certain datasets. Therefore, the current results are insufficient to establish the method's general effectiveness across a broader range of data settings.
Questions for authors

Could the authors clarify the potential conflict between the parameter constraints in the asymptotic consistency theorem and the default parameter settings? Under what conditions can the actual greedy algorithm be guaranteed to be consistent?

Reviewer_Ze3KOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes AQACD, a causal structure learning framework that treats adjacencies and orientations as defeasible hypotheses. It combines conditional-independence evidence, graph-dependent structural arguments, and BIC-based preferences through quantitative argument aggregation. Graph edits are guided by an internal-consistency objective, while orientation decisions additionally require an evidential margin and a risk guard. A central mechanism is structural rescue, which allows previously removed adjacencies to be reconsidered.

The evaluation primarily uses data sampled from five standard discrete Bayesian networks, with 20 independent datasets of 5,000 observations per network. The paper reports improvements on several structural, orientation, and intervention-distance metrics, supplemented by ablations, sample-size experiments, parameter sensitivity analyses, and runtime profiling. It also presents convergence, complexity, and asymptotic-consistency arguments.

Critical strengths

A consequential problem and a concrete revision mechanism.The paper addresses the propagation of finite-sample CI errors from skeleton estimation to orientation. Allowing adjacency decisions to be reconsidered is a meaningful extension over methods that evaluate arguments on a fixed skeleton. The structural-rescue mechanism provides a concrete implementation of this motivation.

Promising empirical results with multiple evaluation perspectives.The paper evaluates structural recovery, orientation, and intervention-distance metrics across repeated trials, and includes component ablations. The gains on CHILD and INSURANCE are particularly encouraging. These results justify further investigation of the proposed revision mechanism, although they do not establish uniform superiority.

Critical weaknesses
  1. The central theoretical guarantees contain unresolved gaps and appear stronger than the supporting analysis.Equation (1) maps independence evidence as τind(p)=max⁡(0,(p−α)/(1−α))\tau_{\mathrm{ind}}(p)=\max(0,(p-\alpha)/(1-\alpha)), whereas the consistency proof subsequently requires pooled independence evidence to converge to 1 for nonadjacent pairs. Under standard asymptotically valid continuous-null tests, a p-value under conditional independence is asymptotically uniform, rather than converging to 1. A vanishing significance threshold therefore does not by itself establish the claimed convergence of the evidence strength. The pooling argument requires an additional justification. There is also an apparent incompatibility between assumption A4, which requires λmax⁡(wsparse,wcoll)<γ\lambda\max(w_{\mathrm{sparse}},w_{\mathrm{coll}})<\gamma, and the reported defaults λ=1\lambda=1, wcoll=1.5w_{\mathrm{coll}}=1.5, given that the evidence lower bound implies γ≤1\gamma\leq1. Finally, Appendix A conditions polynomial complexity on a polynomial revision budget and distinguishes consistency of the global minimizer from correctness of greedy descent, while the contribution statements present stronger guarantees. These issues materially affect the claimed soundness of the method.

  2. The paper does not yet provide a consistent account of the implemented revision procedure or sufficient evidence for its claimed error-correction mechanism.Algorithm 1 starts from a complete graph and allows additions and deletions, whereas Table 4 reports zero deletions and describes a preceding statistical-pruning phase. Proposition 1 states that orientations are irreversible, whereas Table 4 reports orientation reversions. The same CHILD setting also has different AQACD F1 values in Tables 1 and 7 (approximately 0.914 versus 0.888) without an adequate explanation.

These discrepancies obscure which algorithm was evaluated and whether the proofs apply to it. Moreover, counts of rescue operations do not establish that the added edges are correct: graph-derived arguments can also reinforce an erroneous candidate structure. The reported threshold sensitivity and aggregate ablations provide useful evidence, but do not directly establish that the acceptance guards distinguish beneficial from harmful revisions. This matters because reliable correction of cascading errors is a central claim.

Questions for authors
  1. Clarification of the consistency proof: Under what assumptions does the pooled independence strength converge to 1 under the proposed p-value mapping? Please explain how this follows despite the usual asymptotically uniform null distribution of valid p-values.

  2. Clarification of theoretical scope: How can the experimental defaults satisfy assumption A4? If they do not, which theoretical guarantees apply to the reported experiments? Please also reconcile the main-text complexity claim with the revision-budget-dependent bound in Appendix A.

  3. Clarification of the evaluated algorithm: Does the implementation start from a complete graph or a statistically pruned skeleton, and can accepted orientations be reversed? Please reconcile Algorithm 1, Proposition 1, and Table 4, and explain the differing CHILD results in Tables 1 and 7.

  4. Evidence supporting the central mechanism: Do the recorded revision logs distinguish true-edge recoveries from false additions and quantify whether accepted revisions improve or worsen ground-truth structural accuracy? If available, this analysis—and a comparison with the relevant guards disabled—would materially affect my assessment of the claimed protection against error cascades. This request concerns validation of the paper’s central mechanism, rather than an unrelated extension.

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

The paper proposes Abductive Quantitative Argumentation for Causal Discovery (AQACD), which represents adjacencies and orientations as defeasible hypotheses. Conditional independence tests, graph motifs, and BIC differences are transformed into supporting or attacking strengths, aggregated into acceptability scores, and used to revise a graph through skeleton and orientation phases. The paper claims finite termination, polynomial complexity under stated conditions, and asymptotic recovery of the true CPDAG. The method is evaluated on five discrete Bayesian network benchmarks, with additional ablations, runtime analysis, significance tests, and baselines in the appendix.

Critical strengths
  1. The integration of CI tests, local scores, and structural motifs offers a potentially useful way to revisit decisions that would be irreversible in a standard constraint-based pipeline.
  2. The empirical evaluation is comprehensive: it includes five main networks, 20 repeated datasets, ablations, orientation and interventional metrics, sample-size and parameter analyses, runtime profiling, and pairwise tests against QACD.
Critical weaknesses
  1. The novelty and positioning are unclear. The accounts of CPC/MPC, weighted ASP, continuous methods, GES, and ABA-PC are at points inaccurate or incomplete. AQACD reuses ABA-PC's core p-value transformation and a product aggregation that appears in DF-QuAD, without clearly isolating what is new in the representation, aggregation, objective, or revision procedure. Rago et al. also do not appear to support the paper's Noisy-OR attribution or probabilistic interpretation (see question I1).

  2. The method is not specified formally enough to reconstruct. The relations among the hypotheses, graph states, argument tuples, QBAF relations, acceptability scores, and objective are unclear. The conditioning-set pool, triad rescue, structural regulariser, edits, recall step, and stopping conditions are also underspecified, with essential definitions deferred to the appendix. It is therefore difficult to determine precisely which algorithm was implemented and evaluated (see questions I2 and I4-I6).

  3. The statistical and empirical components are not internally aligned. The p-value notation changes between independence and dependence evidence, while transformed p-values and BIC differences are combined without a common calibration. The claimed probabilistic interpretation relies on independence among evidence sources that is not established, and HEH_E and HOH_O overlap although they are treated as alternative graph states. Empirically, the main and appendix CHILD values differ, and the ABA-PC comparison is not like for like (see questions I2-I5 and I7).

  4. Several headline claims exceed the formal support provided. These include strict improvement in global consistency, the likelihood-ratio and Neyman-Pearson interpretation of Eq. 1, strictly polynomial time, non-oscillatory termination, and asymptotic recovery. The relevant theorem statements and assumptions appear only in the appendix and are not connected precisely to the graph states, edits, stopping conditions, or evaluated implementation. Their scope and validity therefore remain difficult to assess (see questions I1, I3, and I6).

Questions for authors

I1. Positioning and novelty

  • P1 — "strictly improves global internal consistency", P5 — "globally coherent", and P6 — "Internal Consistency Score" What formal property is meant by internal consistency or global coherence, how is it measured by Φ(G)\Phi(G), and which method is the comparator for the claimed strict improvement, given that exact conflict resolution methods already target coherent solutions?
  • P1 — "Existing causal discovery methods mainly follow three classical paradigms": What organising dimensions are intended? Continuous optimisation is normally an approach to structure learning, while functional causal models are omitted despite the cited survey.
  • P1 — "treat statistical outcomes as rigid binary clauses": How does this apply to the cited weighted ASP work and to the later acknowledgement that weighted ASP and MaxSAT optimise over conflicting evidence?
  • P1 — "Continuous relaxations provide scalable optimization, yet remain sensitive to distributional assumptions and observational equivalence" and P2 — "continuous approaches ... can be unstable when distinguishing Markov-equivalent configurations in finite samples" Which methods and results support this general characterisation? Observational equivalence is not specific to continuous methods, while assumptions and benchmark pathologies are method dependent.
  • P2 — "CPC and MPC ... discard ambiguous triples entirely": In what sense are the triples or their edges discarded? These methods flag conflicting triples as ambiguous and refrain from orienting them.
  • P2 — "GES ... optimize[s] global graph scores" and P3 — "updates are accepted only when an alternative topology notably improves global consistency" How does this search logic differ from accepting graph edits that improve the objective in GES?
  • P2 — "accommodate conflicting CI outcomes globally" and "symbolically by optimizing qualitative assumptions" and P2 — "To accommodate graded statistical uncertainty" Does this characterisation understate that ABA-PC already optimises over conflicting weighted assumptions and converts p-values into graded strengths? AQACD's active branches in Eq. 1 use the same affine rescaling around α\alpha as ABA-PC's γ(p,α)\gamma(p,\alpha) before treating conditioning sets differently; which distinction is intended here?
  • P2 — "assumption growth" and "lacks continuous topological revision": What do these expressions mean? ABA-PC can recover from an adjacency mistake by excluding lower ranked CI-test results, while AQACD also revises a discrete topology through single edge edits.
  • P2 — "QBAFs ... evaluate continuous argument strengths", P5 — "Noisy-OR gradual semantics—the canonical probabilistic operator satisfying conjunctive independence of epistemic sources", and Appendix P26 — "Justification of Noisy-OR Gradual Semantics vs. DF-QuAD" Rago et al. define the same product aggregation 1−∏i(1−vi)1-\prod_i(1-v_i) as part of DF-QuAD, but do not call it Noisy-OR or give it the stated probabilistic interpretation; DF-QuAD also combines support, attack, and a base score differently from Acc(h)=S(h)(1−A(h))Acc(h)=S(h)(1-A(h)). Which component is being attributed to Rago et al., which is new to AQACD, and how does that attribution align with the explicit contrast between Noisy-OR and DF-QuAD in Appendix D.6?

I2. Formal objects and graph states

  • P3 — "Continuous gradual semantics ... maps heterogeneous evidence to [0,1][0,1]", P4 — "Each h∈Hh\in\mathcal H is assigned an acceptability score", P4 — "saturating S(HE)S(H_E) under Noisy-OR aggregation", and P5 — "Noisy-OR gradual semantics" Figure 1 and Eq. 4 give the generic products S(h)S(h) and A(h)A(h) and the rule Acc(h)=S(h)(1−A(h))Acc(h)=S(h)(1-A(h)). Is S(HE)S(H_E) intended to be S(h)S(h) evaluated at h=HE(X,Y)h=H_E(X,Y)? What exactly is the object whose acceptability is computed—a hypothesis, an argument node, or an active graph state—and how does the stated gradual semantics relate to these equations?
  • P3 — "Inductively-Resolved PDAG" and P5 — "practitioners often require a single representative topology" What edge marks can this object contain, how does it differ from a PDAG or CPDAG, and under what condition is it returned? What supports the practical claim, and how does the objective select this representative from an equivalence class?
  • P4 — "HE(X,Y)H_E(X,Y)" and "HO(X→Y)H_O(X\to Y)" and P6 — "sG(X,Y)s_G(X,Y) maps the active topological state" Are these nested hypotheses, mutually exclusive graph states, or two layers of one state? HO(X→Y)H_O(X\to Y) appears to entail HE(X,Y)H_E(X,Y), but they are later treated as alternative active states; how can sGs_G map an unordered pair to HOH_O without retaining whether its direction is X→YX\to Y or Y→XY\to X?
  • P4 — "P∈{support,attack}P\in\{\text{support},\text{attack}\} is its polarity", P4 — "⟨H,Rs,Ra⟩\langle\mathcal H,R_s,R_a\rangle" and "⟨C,P,τ⟩\langle C,P,\tau\rangle", P5 — "evidential arguments", and P5 — "Sarg(h)S_{arg}(h) and Aarg(h)A_{arg}(h)" Are argument tuples nodes in the QBAF, evidence attached to hypotheses, or a separate representation? How does polarity induce RsR_s and RaR_a, how do those relations construct the support and attack sets, and where is any initial or base strength specified? How are empty sets handled, and how do τ\tau and τ′\tau' enter Eq. 4?

I3. Statistical evidence and pooling

  • P4 — "Following evidence theory (Russo et al., 2024)" and P4 — "Their continuous evidential strengths are mapped" What theory is meant, and what are the source quantities and target arguments or hypotheses of this mapping?
  • P4 — "p(X⊥Y∣Z)p(X\perp Y\mid Z)" and P4 — "p(X̸!⊥Y∣Z)p(X\not\\!\perp Y\mid Z)" Does the first expression always denote the p-value of the CI test with null X⊥Y∣ZX\perp Y\mid Z? If so, is the second the same p-value restricted to p<αp<\alpha, its complement, or the output of a different test?
  • P4 — "Eq. 1 acts as a bounded monotone surrogate for the empirical log-likelihood ratio": What establishes this interpretation and the accompanying Neyman-Pearson claim? The cited ABA-PC work defines the p-value normalisation but does not appear to establish either result.
  • P4 — "Evidence Pooling across Conditioning Sets": Why does Eq. 2 use a maximum for independence and a minimum for dependence, which conditioning sets enter these extrema, and are all of them tested?
  • P2 — "dmaxd_{max}" and P4 — "dmaxd_{max}" and "G(t)G^{(t)}" What does dmaxd_{max} bound in the stated polynomial complexity, and where are the search domain and the relation between the tested conditioning sets and the current graph defined?

I4. Structural rescue and orientation evidence

  • P2 — "dynamic triad rescue", P4 — "un-shielded non-collider", P4 — "unexplained by active paths in the remaining graph G(t)G^{(t)}", and P4 — "τtriad\tau_{triad}" What formally triggers the rescue operation, what is the active-path condition, and how is it checked? Over which triples or conditioning sets is τtriad\tau_{triad} maximised, given that its displayed formula contains neither G(t)G^{(t)} nor an active-path condition?
  • P4 — "preserving the true adjacency": How can the rescue rule know that the candidate adjacency is true, and what prevents it from restoring a false edge?
  • P5 — "Z∉ZsepZ\notin Z_{sep} supports collider" and P5 — "1−p(X⊥Y∣Zsep)1-p(X\perp Y\mid Z_{sep})" Is ZZ the centre of the triple or an element of a conditioning set, and which separating set is used when several have been found? If pp is the CI-test p-value, an accepted independence can have a large value and this strength is then small; is that the intended direction of collider evidence?

I5. Argument aggregation and objective

  • P5 — "ΔBIC(X→Y)\Delta\mathrm{BIC}(X\to Y)" and P5 — "ηN=N\eta_N=\sqrt N" Which two models or parent sets are compared, which sign supports the proposed orientation, and what establishes the N\sqrt N scaling from the cited Chickering result?
  • P4 — "continuous evidential strengths are mapped via piece-wise linear transforms", P5 — "τcoll=min⁡(…,1−p(X⊥Y∣Zsep))\tau_{coll}=\min(\ldots,1-p(X\perp Y\mid Z_{sep}))", P5 — "τbic(X→Y)\tau_{bic}(X\to Y)" and "a unified acceptability score", and Appendix P26 — "synthesizing heterogeneous statistical evidence (p-values) and score variations (τbic\tau_{bic}) through Noisy-OR gradual semantics" How are the strengths derived from p-values aligned with the tanh-normalised BIC strength when both contribute to Acc(HO)Acc(H_O) under Eq. 4? Mapping them to [0,1][0,1] does not make their statistical meanings or magnitudes comparable. What calibrates their relative influence?
  • P4 — "dense chordless cycles without statistical support", P6 — "Ωstruct(G)\Omega_{struct}(G)", P6 — "chordless loop constraints", and P6 — "collider fidelity" How are the cycles on P4 detected and related to the triad argument, given that τsparse=1−τdep∗(X,Y)\tau_{sparse}=1-\tau^*_{dep}(X,Y) does not itself identify a cycle or its statistical support? What are the exact phase-dependent terms for chordless loops and collider fidelity in the regulariser, and is the same objective being optimised throughout the method?
  • P5 — "dialectical margin" and "risk ceiling" and P6 — "the reversal-risk metric" How are the thresholds chosen and justified, and how do A(HO)A(H_O) and 1−Acc(HE)1-Acc(H_E) represent the risk of reversing an orientation?

I6. Algorithm and theoretical claims

  • P1 — "bypassing the exponential bottlenecks typical of exact logical inference", P6 — "at most c≤2c\leq2 state transitions", P6 — "strict Lyapunov descent terminating without cycles", and P6 — "polynomial convergence" Which exact logical formulation and input parameter supply the exponential comparator? Where is the transition bound imposed, which transitions count, and is it what turns finite strict descent into the claimed polynomial result? Does "polynomial convergence" refer to runtime, accepted transitions, or convergence to a particular graph?
  • P2 and P6 — "Theorem 1" and "Theorem 2": The main text mentions the results, but their statements and assumptions are confined to the appendix.
  • P6 — "single-edge edits h∈{+e,−e}h\in\{+e,-e\}" and "G(t)⊕h∗G^{(t)}\oplus h^*": How are candidates generated, and what does ⊕\oplus do for additions and deletions? How are later orientation edits represented?
  • P6 — "Selective Recall": The operation is defined only later in the appendix. How does it interact with Algorithm 1 and evidence accumulated before an adjacency is reconsidered?
  • P6 — "until Φ(G)\Phi(G) converges" and P6 — "until orientations converge" Does the first mean no admissible edit, no score improvement, or numerical stability within a tolerance, and what is the corresponding stopping condition for the orientation phase?

I7. Empirical comparisons

  • P7, Table 1 — "CHILD" and P7, Table 1; Appendix P19, Table 7 — "AQACD" and "QACD" Why is ABA-PC absent when Appendix Table 7 reports a completed CHILD run? Were the same datasets, seeds, settings, and evaluation code used, given that Table 7 reports F1=0.888F_1=0.888 and 0.8870.887, whereas Table 1 reports 0.9140.914 and 0.8380.838? Which results form the like-for-like comparison?
  • P8, Table 2: Are these results significant?
  • P9, Figure 3: Why does the figure omit several baselines reported in Table 1?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.