acceptodds
Under review as a conference paper at ICLR 2027

SCORE-EM: Ordering-Based Causal Discovery with Missing-at-Random Data

Abstract

Causal discovery from observational data is fundamental to machine learning and statistics, yet its application to real-world datasets is often complicated by the prevalence of missing values. Among ordering-based approaches, SCORE (Rolland et al., 2022) estimates a topological ordering using score-function identities under an additive-noise model, but cannot be directly applied to incomplete data. We introduce SCORE-EM, an EM-inspired framework for causal discovery under missing-at-random (MAR) mechanisms that jointly refines missing-value estimates and causal structure ordering. In the conditional-sampling step, missing entries are updated using Metropolis-adjusted Langevin algorithm (MALA) transitions targeting the conditional distribution induced by the current additive-noise model. In the structural-refinement step, the model is refitted and the causal ordering is updated using SCORE statistics aggregated across the resulting ensemble of completed datasets. Unlike structure-agnostic imputation, this procedure uses the evolving structural model to guide missing-value estimation, while the refined imputations provide increasingly informative data for structure learning. A subsequent multiple-imputation CAM pruning stage accounts for imputation uncertainty during edge selection. Empirical results on synthetic and real world data demonstrate that SCORE-EM improves causal ordering and graph recovery over competing missing-data baselines.

Primary area:
causal reasoning
Keywords:
Causal Discovery, Expectation Maximization, Score Matching, Causal Ordering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_vDQKOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SCORE-EM, an EM-inspired framework for causal discovery from missing-at-random (MAR) data. The method addresses the difficulty of applying SCORE to incomplete observations, where conventional imputation may distort the density-score information used for causal ordering. SCORE-EM iteratively updates missing values through MALA-based conditional sampling under a fitted additive noise model (ANM), followed by structural model refitting and causal ordering refinement using pooled SCORE statistics. The final DAG is reconstructed using multiple-imputation CAM pruning.

The authors evaluate SCORE-EM on synthetic nonlinear ANMs with Erdős–Rényi and scale-free graph structures at 30% and 50% missingness, as well as on the Sachs protein-signaling dataset with artificially introduced MAR missingness. The proposed method achieves higher mean F1 scores than the evaluated missing-data baselines across the synthetic configurations and consistently improves mean order divergence over other ordering-based methods. Additional statistical tests support the aggregate performance improvements, although the relative performance in SHD varies across graph structures and missingness settings.

Critical strengths
  1. A well-motivated approach to causal discovery with incomplete data.

The paper addresses an important and practically relevant limitation of SCORE, which requires complete observations to estimate the density-score quantities used for causal ordering. The motivation is clearly presented, and the proposed framework provides a concrete approach to incorporating structural information into missing-value estimation.

In particular, Appendix 13 offers a useful mathematical illustration of how point imputation and misspecified nonlinear multiple imputation can distort SCORE's Hessian-based leaf criterion. Although this example does not establish the superiority of SCORE-EM over correctly specified multiple-imputation methods, it provides a sound motivation for considering structure-aware conditional sampling.

  1. Encouraging empirical improvements in causal ordering and graph recovery.

SCORE-EM achieves the highest reported mean F1 among the evaluated missing-data methods across all 24 synthetic configurations and consistently obtains lower mean order divergence than the other ordering-based baselines. Appendix 12 further reports statistically significant aggregate improvements using paired Wilcoxon signed-rank tests with Holm correction.

These results provide meaningful empirical evidence for the effectiveness of the overall SCORE-EM framework. In particular, the consistent improvements in order divergence are relevant because accurate causal ordering is itself an important objective of ordering-based causal discovery.

Critical weaknesses
  1. Insufficient characterization of the mathematical scope and practical robustness of end-to-end DAG recovery.

MAR ignorability (Section 3.2) does not by itself guarantee causal identifiability. The manuscript does not establish sufficient joint-observation, positivity, and structural assumptions under which the full-data distribution and causal ordering or DAG can be identified from MAR observations at the population level. Nor does it provide theoretical results characterizing how recovery guarantees or estimation error deteriorate as missingness increases. Experiments at 30% and 50% missingness under specific MAR mechanisms do not adequately characterize the conditions under which reliable recovery can be expected.

Furthermore, Equation (1) allows general nonlinear parent interactions, whereas provisional parent selection (Appendix 9.4) and final CAM pruning (Appendix 10) use additive regression models. For example, with independent standard Gaussian root variables, the interaction-only mechanism X3=(X12−1)(X22−1)+Z3X_3=(X_1^2-1)(X_2^2-1)+Z_3 contains genuine causal parents whose population additive effects vanish. Thus, correct causal ordering need not guarantee correct edge recovery. Although this limitation is also inherited from CAM-based approaches, the applicable functional assumptions should be clarified.

The experimental results further illustrate the importance of characterizing end-to-end recovery: at 50% missingness, MissDAG achieves lower mean SHD in all six sparse ER1/SF1 settings (Appendix Tables 6–8), despite SCORE-EM's favorable F1 and ordering results among the relevant baselines. These setting-specific exceptions do not negate the positive aggregate results, but they leave the reliability of final DAG recovery insufficiently understood.

  1. Insufficient evidence isolating the contribution of iterative structural refinement.

The central algorithmic distinction is the iterative feedback between ANM-guided imputation and SCORE-based ordering, although MissDAG already couples missing-data inference with structural learning. Section 6.1 treats MI+SCORE as an ablation baseline, but the methods differ in their imputation models (Extra-Trees versus ANM-guided MALA), ensemble sizes (20 versus 4), and pruning-stage imputations (Appendices 8–10).

Consequently, the reported improvements support the effectiveness of the overall SCORE-EM pipeline but do not isolate the contribution of iterative structural feedback from these other design choices. This limits the evidence for the paper's distinctive algorithmic contribution and its significance.

Questions for authors
  1. Population identifiability and theoretical recovery under MAR

Under what joint-observation patterns, positivity conditions, and nonlinear ANM assumptions can causal ordering or DAG structure be identified from MAR observations at the population level? Under such conditions, can the authors establish theoretical recovery guarantees or estimation error bounds characterizing the deterioration in performance as observation probabilities decrease, including their dependence on sample size and signal strength?

  1. Assessing identifiability conditions in practice

How could practitioners assess whether their incomplete data plausibly satisfy the model assumptions and identification conditions needed for causal recovery under SCORE-EM? In particular, which conditions can be assessed from observable information (such as joint-observation overlap, model diagnostics, or stability of estimated SCORE leaf statistics), and which cannot be verified without additional assumptions? How should practitioners assess sensitivity when these conditions cannot be established from the data alone? 3. Functional assumptions and CAM pruning

What is the intended functional model class for end-to-end DAG recovery, given that Equation (1) allows general nonlinear interactions while the parent-selection and CAM pruning procedures use additive regressions? In particular, how does the method account for interaction-only parental effects that have zero population additive contributions?

  1. Independent contribution of iterative refinement

Can the authors provide a controlled ablation comparing full SCORE-EM with a version that removes iterative structural or ordering feedback while keeping the imputation model, number of completions, pruning procedure, and computational budget as comparable as possible? This would help determine the independent benefit of the proposed feedback mechanism.

  1. Population soundness of the iterative procedure

Assuming exact density-score estimation and exact conditional sampling from the fitted ANM, does the population-level SCORE-EM iteration recover the true causal ordering? Can an incorrect ordering and the imputations fitted under it form a self-consistent fixed point? If so, what guarantee does the proposed procedure provide?

3 comments

We thank the reviewer for the thoughtful comments.

W1–W2. MissDAG's lower SHD on sparse graphs at 50% missingness arises in pruning, not ordering, and under identical pruning SCORE-EM's ordering lowers OD from 1.73 to 0.58 relative to MI+SCORE; both ablations are in our response to Reviewer yeq6.

Q1. One sufficient setting is a fully observed subset ZZ and partially observed variables WW, with RR indicating joint observation of all WW. If R⊥W∣ZR\perp W\mid Z and P(R=1∣Z=z)>0P(R=1\mid Z=z)>0 almost surely, then p(z,w)=p(z)p(w∣z,R=1)p(z,w)=p(z)p(w\mid z,R=1) identifies the joint distribution; our synthetic masks have this form. More generally, if each missingness pattern's probability depends only on its observed entries and complete rows have positive probability, the ignorable likelihood (Eq. 3) identifies the joint law; this covers Sachs.

Given the joint distribution, the DAG is identified under causal sufficiency, acyclicity, and smooth nonlinear ANMs with Gaussian noise (Peters et al., 2014), and population SCORE recovers a valid order (Rolland et al., 2022, Lemma 1). Let γ>0\gamma>0 be the minimum nonleaf Hessian-diagonal variance over all valid leaf-removal sequences. If estimated variances are within γ/2\gamma/2 of their oracle values at every stage, recursive SCORE recovers a valid order. This does not show that SCORE-EM attains this accuracy, and our estimator's DAG recovery also requires consistent parent selection (Q3).

Imputation bias vanishes at the true model for any missingness rate, so the rate enters through estimation error; bounds in observation probabilities, sample size, and signal strength require controlling score-estimation, imputation, and refinement errors jointly, which we leave to future work.

Q2. Practitioners can inspect complete-case fractions, pairwise co-observation counts, and joint counts for each node with its estimated parents. Model-dependent diagnostics include sampler stability, residual normality and constant variance, held-out predictive coverage, leaf-statistic gaps relative to between-imputation variability, and edge-selection frequencies. Stability does not establish correctness (Q5).

Observed data cannot generally distinguish MAR from an observationally equivalent MNAR model, establish causal sufficiency, or validate mechanisms outside observed support. We recommend shifting imputed XjX_j values by δσ^j\delta\hat\sigma_j for δ∈[−0.5,−0.25,0,0.25,0.5]\delta\in[-0.5,-0.25,0,0.25,0.5], repeating ordering and pruning, and reporting sensitivity.

Q3. Equation (1) is broader than the class supported by our edge selection. SCORE's population ordering criterion permits interactions, but provisional screening and final CAM pruning use additive regressions and can miss interaction-only parents even with complete data and the true order.

For independent standard Gaussian roots, let X3=(X12−1)(X22−1)+εX_3=(X_1^2-1)(X_2^2-1)+\varepsilon with centered noise. Then E[X3∣X1]=E[X3∣X2]=0\mathbb E[X_3\mid X_1]=\mathbb E[X_3\mid X_2]=0, so the population additive projection is zero. We will clarify in Section 3 and Appendices 9.4 and 10 that edge recovery requires a detectable additive component for each parent and consistent selection. We will add this limitation and the counterexample.

Q4. We compared full SCORE-EM with a variant holding the initial order fixed during MALA sampling and ANM refitting. Both used the same imputation model, number of completions, iteration cap, and MI-CAM pruning, and both estimated a final order from the completed datasets.

Across experiments at d=10d=10, feedback improved mean F1 from 0.746 to 0.759, reduced SHD from 8.85 to 8.58, and reduced OD from 0.950 to 0.688 (28%). The fixed-order variant's OD is already well below MI+SCORE's (1.1–7.0 in Tables 4 and 6), so ANM-guided sampling provides most of the ordering gain and feedback adds the remainder; graph-recovery gains are modest.

Q5. Exact scores and conditional sampling do not guarantee recovery from arbitrary initialization. For independent X,ε∼N(0,1)X,\varepsilon\sim N(0,1) and Y=X2−1+εY=X^2-1+\varepsilon, let XX be MCAR at rate κ\kappa. Under reversed order (Y,X)(Y,X) with fitted X∣Y∼N(0,1)X\mid Y\sim N(0,1), completion yields qκ=(1−κ)pXY+κpXpYq_\kappa=(1-\kappa)p_{XY}+\kappa p_Xp_Y. Under qκq_\kappa, symmetry gives E[X∣Y]=0\mathbb E[X\mid Y]=0 and Var⁡(X)=1\operatorname{Var}(X)=1, so Gaussian regression refitting preserves this conditional. As κ→1\kappa\to1, the curvature variance for XX tends to zero while that for YY stays positive, so SCORE retains the reversed order. Numerically, this requires κ>0.94\kappa>0.94 (standardized) or 0.980.98 (raw) in this example; at κ≤0.9\kappa\le0.9 SCORE rejects the reversed order.

Under MAR, true conditional imputations preserve the joint law, so the truth is also a fixed point. Q1's accuracy condition guarantees the ordering, but not that the iteration reaches it. We claim neither global convergence nor monotone observed-data likelihood, and will clarify these limits and add the counterexample.

Thank you for the thoughtful responses and the additional experiments. The rebuttal clarifies why additive pruning can miss parents whose effects arise only through interactions. My main remaining question is under what model conditions the method can recover the DAG.

Given the true ordering and imputation from the true conditional distribution, let Δjk\Delta_{jk} denote the increase in population prediction error when predecessor kk is removed from the additive regression for node jj. Could the authors state concrete sufficient conditions under which Δjk>0\Delta_{jk} > 0 for every true parent and Δjk=0\Delta_{jk} = 0 for every nonparent? A correctly specified additive model with nonredundant parent effects would be a natural starting point. Stating these conditions would clarify when the edge-selection step can be expected to work.

Thank you for this suggestion. A concrete sufficient setting is a correctly specified additive SEM,

\qquad \varepsilon_j\overset{\mathrm{ind}}{\sim}N(0,\sigma_j^2),\quad \sigma_j^2>0,$$ where each true-parent function is continuous, nonconstant, square-integrable, and belongs to the additive regression class used for population pruning. We assume the regression class contains the true conditional mean. Let $P_j$ denote all predecessors under the true topological ordering. Under MAR, exact sampling of missing entries from the true conditional distribution preserves the full-data distribution at the population level. Since the structural noise is independent of all predecessors, the optimal population squared prediction error using the full additive model is $\sigma_j^2$. Removing a nonparent leaves the true conditional mean representable, so $\Delta_{jk}=0$. For every true parent, $$\Delta_{jk}\ge \mathbb E\left[ \mathrm{Var}\left( f_{jk}(X_k)\mid X_{P_j\setminus\lbrace k\rbrace} \right) \right]>0.$$ The lower bound follows because even unrestricted prediction from the remaining predecessors cannot eliminate this conditional variation. Strict positivity follows from the everywhere-positive joint density implied by independent nondegenerate Gaussian noises, together with the continuity and nonconstancy of each parent function, which exclude deterministic redundancy. Thus, under these conditions, population risk-based pruning recovers exactly the true parent sets, conditional on the true ordering and oracle imputation. This result characterizes the population separation underlying CAM pruning; consistency of finite-sample spline estimation, significance testing, and imputation requires additional conditions. We will add these sufficient conditions and the corresponding argument to clarify the scope of DAG recovery.
Reviewer_LHvzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies causal structure discovery from data with MAR missingness. SCORE-EM alternates between missing-value estimation and causal ordering: MALA updates missing entries under the current additive-noise model, pooled SCORE statistics update the ordering, and multiple-imputation CAM pruning selects the final edges. Experiments use synthetic graphs with 10 to 30 nodes and the Sachs dataset with artificially introduced missingness. SCORE-EM outperforms two-stage imputation and other missing-data methods in several synthetic settings. On Sachs with 50% missingness, its F1 is 0.352 versus 0.302 for MI+SCORE. At 30% missingness, it matches the F1 of MissForest+SCORE.

Critical strengths
  1. Combining structure-conditioned sampling with SCORE ordering addresses how generic imputation can distort leaf statistics. The method connects missing-value estimation to structure discovery rather than simply substituting an imputer.
  2. Across graph structures, densities, and missingness rates, the method improves ordering and graph recovery over two-stage and other missing-data baselines, supporting the value of this combination.
Critical weaknesses
  1. The controls do not isolate structure-conditioned sampling from repeated ordering updates, and the evaluation does not adequately examine whether an incorrect initial ordering reinforces itself through imputation. This directly affects the assessment of the necessity and stability of the central feedback mechanism.
  2. Real-data validation uses only Sachs with artificial missingness generated from the true causal graph. Without comparisons across MAR mechanisms on the same data, the sensitivity of the recovery advantage to the masking mechanism remains unclear.
Questions for authors
  1. Have fixed initial orderings been compared with repeated ordering updates? Across initializations, which ordering errors does the joint procedure correct, and when does it retain or reinforce them?
  2. Does the ordering and graph-recovery advantage on Sachs persist under the synthetic experiments’ MAR scheme based on randomly selected fully observed variables?
2 comments

We thank the reviewer for their thorough evaluation and insightful feedback.

Q1. We ran additional experiments as per your suggestion.

Fixed initial ordering versus repeated ordering updates. Repeated ordering updates improve causal ordering accuracy, reducing mean order divergence by 28%, while yielding smaller but consistent improvements in graph recovery. Enabling order revision reduced mean OD from 0.950 to 0.688, raised F1 from 0.746 to 0.759, and reduced SHD from 8.85 to 8.58. These results demonstrate the independent benefit of iterative ordering feedback, particularly for ordering accuracy. To isolate the contribution of iterative ordering feedback, we compared full SCORE-EM with a variant that holds the initial order fixed. Both variants perform structure-conditioned MALA sampling and refit the additive-noise model with the same imputation model, number of completions, iteration cap, and MI-CAM pruning. Full SCORE-EM additionally uses pooled SCORE statistics from updated imputations to revise the order during iterative refinement. Results cover 160 paired runs at d=10d=10.

MetricFixed-order SCORE-EMFull SCORE-EM
Order divergence (OD) ↓0.950 ± 0.9800.688 ± 0.836
F1 ↑0.746 ± 0.1380.759 ± 0.134
SHD ↓8.85 ± 4.918.58 ± 4.67

Sensitivity to initialization. We examined ordering correction under the likelihood safeguard (Algorithm 1), measuring OD on the last accepted order before the final pooled recomputation. We compared the natural initialization with controlled starts containing exactly five ordering violations, across 20 datasets at two MAR rates (40 runs per initialization).

InitializationMean initial ODMean final ODDecreased / unchanged / increased
Natural2.552.407 / 31 / 2
Perturbed (OD 5)5.002.7025 / 14 / 1

Which errors are corrected or retained. Correction of perturbed starts is concentrated in sparse graphs: mean final OD fell to 0.7 on ER1 and 1.5 on SF1, compared with 4.1 on ER4 and 4.5 on SF4. On dense graphs, most injected violations therefore persist. Natural initializations were largely unchanged (31 of 40 runs), which indicates that the procedure corrects gross ordering errors on sparse graphs but rarely removes the violations remaining after a natural start.

Self-reinforcement. OD increased in 1 of 40 perturbed runs and 2 of 40 natural runs, so worsening is uncommon but does occur. We do not claim that the procedure cannot retain an incorrect ordering: as shown in our response to Reviewer vDQK (Q5), an incorrect ordering and the imputations fitted under it can form a self-consistent fixed point at the population level. Stability of the estimated order is therefore not evidence of its correctness, and we will state this limitation, together with the dense-graph result above, in the revised manuscript.

Q2. Theoretically, under MAR with distinct parameters, the missingness mechanism is ignorable, and the MALA target p(X_mis | X_obs; θ) does not depend on the specific form of p(R | X_obs). SCORE-EM therefore requires no mechanism-specific correction, unlike propensity-weighted approaches such as MissScore.

We ran empirical experiments with a different MAR mechanism with beta=2 and nonlinear logistic mask on randomly selected fully observed variables for 30% missingness over 3 seeds. The results are as follows:

MethodF1 ↑Precision ↑Recall ↑SHD ↓OD ↓
SCORE-EM0.434 ± 0.0090.564 ± 0.0310.353 ± 0.00012.67 ± 0.587.67 ± 0.58
MissForest+SCORE0.419 ± 0.0160.517 ± 0.0480.353 ± 0.00013.67 ± 1.158.00 ± 0.00
MI+SCORE0.424 ± 0.0090.530 ± 0.0260.353 ± 0.00013.33 ± 0.588.00 ± 0.00
MissDAG0.272 ± 0.0840.606 ± 0.1830.176 ± 0.05914.00 ± 1.00—
MissScore0.224 ± 0.0460.220 ± 0.0600.235 ± 0.05924.00 ± 3.618.67 ± 4.51
MissOTM0.314 ± 0.0470.579 ± 0.0840.216 ± 0.03413.67 ± 1.15—

Thank you for your rebuttal. I choose to keep my original score.

Reviewer_yeq6OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses ordering-based causal discovery when data are missing at random (MAR). The authors argue that generic imputation distorts SCORE's leaf criterion: point imputation shrinks residual variability, and misspecified multiple imputation creates mixture residuals whose Hessian diagonal has nonzero variance (illustrated in Appendix 13). SCORE-EM addresses this with an EM-inspired loop over an ensemble of M completed datasets: Iterations stop based on a Stein-Hessian drift and between-imputation diagnostic. Edges are then pruned by running CAM's partial F-tests on multiple Bayesian-ridge imputations and keeping each edge by stability voting. The method is evaluated on GP-based ANMs (ER/SF graphs, d ∈ {10, 20, 30}, 30% and 50% missingness) and on the Sachs dataset with synthetic MAR masking. It is compared against MissScore, MissDAG, MissOTM, MissForest+SCORE and MI+SCORE, and Wilcoxon signed-rank tests over 480 paired evaluations are reported.

Critical strengths

1.Missing data is common in practice, and ordering-based methods such as SCORE have no principled way to handle it. The paper targets a real gap. 2. The three-variable example in Appendix 13 shows concretely how point imputation and multiple imputation each break SCORE's constant-curvature property for leaves. This is a useful observation in its own right. 3. Updating missing values from the conditional distribution of the current structural model, rather than with a structure-agnostic imputer, is a natural idea. MALA fits it well, since it needs the target only up to a normalizing constant.

Critical weaknesses
  1. The contribution is incremental. Coupling model-based imputation with structure learning via EM is the idea behind MissDAG; here its M-step is replaced by SCORE ordering. As the authors acknowledge, the procedure has no common objective and no convergence or consistency result.
  2. The ordering is often near-oracle, but the graph is not: on SF1 with d=20, the OD equals that of fully observed SCORE, yet the SHD is 8.6 vs 1.9. Pruning uses a separate linear Bayesian-ridge imputer instead of the EM samples, which goes against the paper's own motivation. 3, At 50% missingness on sparse graphs, the SHD exceeds the number of true edges, so it is worse than an empty graph.
  3. MI+SCORE serves as the ablation for the ordering stage, but its pruning uses a different imputer (an Extra-Trees ensemble rather than Bayesian ridge). The comparison is therefore not clean.
  4. The scope is narrow: Gaussian noise only, d ≤ 30, and a single MAR mechanism. Only an asymptotic complexity analysis is given; no wall-clock runtimes are reported, although each iteration runs M imputations with L MALA transitions.
Questions for authors
  1. What happens if pruning uses the EM completions? Could you run a cross ablation, swapping the ordering and pruning stages between SCORE-EM and MI+SCORE? How many seeds are used per setting? The main tables do not say, while Table 3 pools N=480 paired evaluations. For several ordering-based baselines, the OD is close to the other methods but the SHD is far larger, sometimes several times the number of true edges. This points to the pruning or thresholding stage. Could you report the pruning procedure and thresholds used for each baseline?
2 comments

We thank the reviewer for the constructive feedback.

W1. SCORE-EM is not MissDAG with a different M-step. MissDAG maximizes a likelihood under a continuous acyclicity constraint; SCORE-EM updates a discrete ordering from Hessian-diagonal statistics pooled over completions, with no shared objective, which is why pooled leaf statistics and a likelihood safeguard are required. We position this as a focused extension without a global convergence guarantee. SCORE-EM attains higher mean F1 than MissDAG in all 24 settings, lower mean SHD in 18, and lower runtime (see W4).

W2. SHD–OD gap. Correct. Pruning under missingness is the bottleneck, even with the true ordering. On ten d=10d=10 ER1 datasets, we crossed ground-truth/SCORE-EM ordering with complete/imputed data for CAM pruning (SHD below). With imputed data, SCORE-EM's ordering matches the true ordering (9.7 vs 9.6 at 50%), so ordering is not the limiting factor. We will add this ablation and state the limitation.

OrderingPruning30%50%
Ground truthComplete0.80.8
Ground truthImputed3.09.6
SCORE-EMComplete1.61.7
SCORE-EMImputed (ours)2.89.7

Pruning on EM completions. On d=10d=10, n=1000n=1000 ER/SF data with mask, order, CAM tests and voting threshold fixed, separate MI gives higher F1 overall and on dense graphs; EM completions are better on sparse graphs at 50%. EM completions are drawn under the screened ANM, so a true parent excluded by the screen is imputed without its effect on the child, biasing its test toward the null. We keep separate MI as the default and will report EM pruning as a variant for sparse graphs.

F1EM, 30%MI, 30%EM, 50%MI, 50%
All0.7590.8070.6570.689
Sparse0.8040.8000.7240.618
Dense0.7140.8140.5900.759

SHD exceeding true edges. Correct: at 50% with d=20d=20, SHD (28.0 ER1, 25.4 SF1) exceeds the true edge count (18.7, 19.0). Unlike an empty graph, SCORE-EM recovers 17.2 and 15.8 of these edges (TPR 0.917/0.832) and rejects about 93% of ordered non-edges (FPR 0.074/0.065). With about 19 ordered non-edges per true edge, a 7% FPR yields more spurious than true edges; precision is about 0.40. The output is a high-recall superset of the true graph; MissDAG and MissOTM have lower SHD here, at lower F1.

W3. We agree. Appendix 10 misstates our configuration: the reported SCORE-EM results use Extra-Trees pruning, as MI+SCORE does (not Bayesian ridge); we will correct it. We also crossed MI/SCORE-EM orderings with Bayesian-ridge (BR)/Extra-Trees (ET) pruning under identical settings on d=10d=10 ER1/ER4/SF1/SF4 graphs at 30% missingness (40 datasets). OD falls from 1.73 to 0.58, independent of pruning by construction. With the imputer fixed, SCORE-EM improves F1 by 0.050 (BR) and 0.084 (ET) and reduces SHD by 0.73 and 1.85. We will include this ablation.

OrderingImputerF1 ↑SHD ↓
MIBR0.724±0.1328.83±6.98
MIET0.735±0.1128.20±4.73
SCORE-EMBR0.774±0.1518.10±7.55
SCORE-EMET0.819±0.1096.35±4.82

W4. We measured wall-clock runtimes. SCORE-EM takes 97 ± 13 s, 344 ± 29 s and 699 ± 100 s at d=10d=10, 2020 and 3030, and attains the highest F1 at every dd. Relative to MI+SCORE, it is 1.95× and 1.20× faster at d=10d=10 and 2020, and 11% slower at d=30d=30, with F1 higher by 0.097–0.127. It is faster than MissDAG at every dd. MissForest+SCORE is faster (roughly 7× at d=10d=10, 3× at d=30d=30), at an F1 cost of 0.068–0.099. Since the learned graph is the basis for subsequent scientific analysis, we consider a cost of minutes justified by the consistent F1 gain. We acknowledge that the synthetic study is limited to Gaussian noise and d≤30d\leq30, and will state this as a limitation. On the MAR mechanism, the Sachs experiment already uses a parent-dependent mechanism distinct from the synthetic one (Section 5.2), and we have run an additional mechanism on Sachs (see our response to Reviewer LHvz).

Questions.

EM completions: W2. Cross ablation: W3.

Seeds. The main tables report 10 datasets per setting at n=1000n=1000. Table 3 pools these with 10 further datasets per setting at n=2000n=2000 (24 settings × 20 = 480). We will state this in Section 5.1 and also report Table 3 restricted to n=1000n=1000.

Pruning and thresholds. The pattern is driven mainly by MissScore: on SF1, d=20d=20, 30% missingness, its OD (3.4) is close to MI+SCORE's (3.2), yet its SHD is 43.0 versus 18.0 (about 19 true edges), so the gap arises in edge selection. All ordering-based methods prune at significance 10−310^{-3}. SCORE-EM and MI+SCORE apply spline partial F-tests to Extra-Trees multiple imputations, keeping edges significant in at least half of the completions; MissForest+SCORE applies CAM pruning to one MissForest completion; MissScore uses its authors' modification of CAM-style pruning.

Thank you for the detailed response. My main concern is that the work seems somewhat incremental, essentially combining score matching with EM. I don't see a clear methodological/theoretical insight in the proposed approach. I keep my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.