acceptodds
Under review as a conference paper at ICLR 2027

Discrete Score Matching Enables Causal Discovery from Count Data

Abstract

Count data pose a challenge for score-matching-based causal discovery: derivatives are unavailable, and simply replacing them with finite differences does not generally suffice for causal discovery. We generalize SCORE's constant-curvature criterion (Rolland et al., 2022) by conditioning on the node's value, yielding the conditional curvature score (CCS) for ordering. We also extend curvature-based parent recovery through the off-diagonal curvature score (OCS), enabling directed acyclic graph (DAG) recovery with both scores constructed from score functions for continuous data and concrete scores for counts. In the bivariate setting, zero CCS exactly characterizes a semiparametric generalized linear model (GLM) conditional form in which the conditional family need not be specified in advance, unlike in classical GLMs. For bivariate semiparametric GLM DAGs under our regularity condition, canonical-parameter nonlinearity is necessary and sufficient for identifiability. In multivariate DAGs, this nonlinearity enables DAG recovery through CCS and OCS. Our framework identifies a new class of semiparametric GLM DAGs that strictly contains the nonlinear Gaussian ANM class identified by SCORE. We introduce DISCO (DIscrete SCOre), a count-DAG recovery algorithm that estimates CCS and OCS using discrete diffusion. Experiments demonstrate accurate DAG recovery across Poisson, negative binomial, binomial, and mixed-family settings, as well as scalability to 1,000-node DAGs on a single GPU.

Primary area:
causal reasoning
Keywords:
causal discovery, score matching, discrete diffusion, count data

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_YwneOpenReview ↗
Rating
3
Confidence
3(+1)
Informativeness
3
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper extends score-matching-based causal discovery to count data. SCORE's constant-curvature sink criterion fails for counts because the node's own conditional density contributes an ``own-factor'' curvature that varies with the node's value even at a sink. The paper introduces the conditional curvature score (CCS)---the conditional variance of the diagonal log-density curvature given the node's value---which absorbs own-factor variation while retaining child-induced variation, and uses it to order nodes; an off-diagonal curvature score (OCS) then recovers parents given the order. Both scores are computed from concrete scores estimated by score-entropy discrete diffusion. The theoretical results are: in the bivariate setting, zero CCS exactly characterizes a semiparametric GLM conditional form whose family need not be prespecified; within this class, canonical-parameter nonlinearity is necessary and sufficient for identifiability; multivariate DAGs are recoverable by CCS ordering plus OCS parent selection; and a plug-in consistency result holds under an assumed convergence rate for the curvature estimates. The identifiable class strictly contains SCORE's nonlinear Gaussian ANMs. Experiments show that DISCO outperforms count-DAG baselines at scale on synthetic data, with a Lahman batting analysis as a real-data diagnostic.

Critical strengths
  1. The model-class relaxation is well motivated and correctly sized relative to prior work. Existing routes to causal discovery from counts either do not cover count support at all (SCORE's nonlinear Gaussian ANMs) or require the conditional families to be prespecified: ODS (Park and Raskutti) within Poisson/QVF families, MRS (Park and Park) within generalized hypergeometric families, and CPCM (Bod'ik and Chavez-Demoulin) within a fixed list of parametric families. This paper drops the prespecified family while keeping a one-parameter exponential-tilt structure, so the conditional families need not be known in advance. The resulting class strictly contains SCORE's nonlinear Gaussian ANMs (the Gamma example: nonconstant sink curvature yet zero CCS) and covers the negative-binomial and binomial settings that motivate the experiments.

  2. On the algorithmic side, the CCS construction is the right fix to a precisely diagnosed obstruction. SCORE's constant-curvature criterion fails on counts because the node's own conditional density contributes curvature that varies with the node's value even at a sink; prior count-DAG methods respond by presetting families and using moment conditions instead. Conditioning on the node's value to absorb own-factor curvature is simple, effective, and cleanly unified---one coordinate operator treats continuous and count coordinates in a single framework---and the rank-transform invariance (Proposition 12) is a practically valuable property that the experiments confirm. The ablations isolate the gain over the constant-curvature criterion (+13.4+13.4/+20.0+20.0 sink-selection points on Poisson/NB at d=50d = 50, n=5000n = 5000), and the scalability is demonstrated concretely (d=1000d = 1000, median F1=0.769F_1 = 0.769 in 57.4 minutes on a single GPU; Table 3).

  3. The theory is coherent and, as far as I checked, correct under the stated assumptions. The bivariate necessary-and-sufficient characterization (Theorem 6), the multivariate recovery theorem (Theorem 8), and the strict-extension result (Corollary 13) fit together as one story, and I did not find substantive errors in the parts of the proofs I worked through. The limitations are scoped with unusual honesty (e.g., Appendix E.2's explicit statement of what is not proven).

Critical weaknesses
  1. The estimation theory does not cover the implemented algorithm---a gap the authors themselves acknowledge. Assumption 15 assumes the uniform convergence of the curvature estimates, which is precisely what an end-to-end guarantee would need to prove from the score-entropy estimator [1], the amortized projection, the rank transform, and the boundary clipping; no convergence rates are available at any stage; the analyzed variant (raw OCS thresholding at a deterministic λn\lambda_n) differs from the implementation (boundary clipping, median/IQR standardization, selection-frequency rule); and λn\lambda_n must be sandwiched between a rateless error level and the smallest true-edge signal, so no explicit choice is available. The authors state this openly (Appendix E.2: "does not give a global consistency guarantee''), and I appreciate the honesty; for a paper whose contribution is primarily theoretical, this remains the main incompleteness.

  2. The closest prior method is not experimentally compared. Vo et al. (2026) [2]---ordering-based causal discovery via generalized score matching for discrete data---is the nearest relative of DISCO's approach and is discussed in the related-work section, yet it appears in no experiment. Without a head-to-head comparison, the empirical assessment leaves open how the two methods' coverage differs (which conditional families and data regimes each supports), which is better matched to real count data in practice, and how they perform on the same benchmarks. The comparisons that are present also carry caveats: the multivariate ODS baseline [3] uses regression estimates of the conditional mean and variance rather than the original exact-conditioning construction, so the Oracle gap may partly reflect this adaptation.

[1] A. Lou, C. Meng, and S. Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution. ICML, 2024.

[2] V. Vo, T. Le, H. Zhao, E.,V. Bonilla, and D. Phung. Ordering-based causal discovery via generalized score matching. KDD, pages 4706--4717, 2026.

[3] G. Park and G. Raskutti. Learning quadratic variance function (QVF) DAG models via overdispersion scoring (ODS). Journal of Machine Learning Research, 18(224):1--44, 2018.

Questions for authors
  1. The conditional form characterized by Theorem 5 seems to me related to the classical semiparametric density-ratio literature [1,2], and the affine-case non-identifiability in Theorem 6 to the symmetry behind reversible linear logistic models [3]. Is this reading correct? If so, how does the paper's contribution differ from what this literature establishes? Although DISCO's assumptions are more general, my sense is that its identifiability mechanism shares its core algebraic step with CPCM [4], and that neither class fully covers the other. Could you explain in detail the similarities and differences of the identifiability mechanisms, and state the cross-coverage explicitly?

  2. My reading of the proof of Theorem 8(i) (Appendix C.3) is that it uses only the existence, for each nonsink jj, of a sink child in the induced child subgraph whose edge is nonlinear---strictly weaker than Definition 4's every-edge nonlinearity. Is this correct, and if so could the ordering theorem be stated under the weaker per-node condition? My reading of Condition R1 is that it supplies, at count-support endpoints, the interiority that the squeezing argument already yields at interior levels. Is this correct, and could you provide a concrete endpoint example? More generally, would you surface the regularity conditions in the main text, together with a discussion of what they exclude (parent-dependent dispersion, zero inflation, near-deterministic relations)?

  3. Under what explicit conditions could Assumption 15 be verified---for example, at a single fixed noise level with exact marginalization, importing a consistency rate for the score-entropy estimator from Lou et al.\ (2024) [5]? Would you state any stage-wise rate, even under strong conditions?

  4. The theory is established on the admissible finite-difference domain, under which the Lahman headline is 0.762 (stable across q4--q8), while the reported 0.905 uses the clipped boundary convention. What considerations favor the clipped convention on real data? Relatedly: the seven reference relations are accounting containments such as H≤ABH \leq AB---the set of values a child can take depends on its parent's value---whereas Definition 3 requires each node's support and density to be independent of the parent values (strict positivity on a fixed support). The data-generating process therefore appears to lie outside the model class in which the identifiability guarantees hold. Given this, would you consider presenting the Lahman section explicitly as a diagnostic (agreement with domain knowledge) rather than as evidence of causal recovery?

[1] J.,A. Anderson. Multivariate logistic compounds. Biometrika, 66(1):17--26, 1979.

[2] J. Qin. Inferences for case-control and semiparametric two-sample density ratio models. Biometrika, 85(3):619--630, 1998.

[3] R.,L. Prentice and R. Pyke. Logistic disease incidence models and case-control studies. Biometrika, 66(3):403--411, 1979.

[4] J. Bod'ik and V. Chavez-Demoulin. Identifiability of causal graphs under non-additive conditionally parametric causal models. Journal of Machine Learning Research, 26(264):1--55, 2025.

[5] A. Lou, C. Meng, and S. Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution. In ICML, 2024.

6 comments

We are thankful for the reviewer's close reading of the proofs and the questions on the estimation theory, related literature, and real data. Through them, the revised manuscript now gives stage-wise rates (Propositions 11A and 15A) and clarifies how DISCO relates to existing methods and how the real-data results should be read. More importantly, the reviewer's reading of the proof of Theorem 8(i) led to Theorem 9A, which allows linear edges in count DAGs. Revised passages of the manuscript are marked in blue.

W1. Gap between the estimation theory and the implemented algorithm

The boundary convention now follows the theory, Assumption 15 can be verified with explicit rates for the empirical score-entropy minimizer on a finite support, and the standardized OCS rule is one scalable choice that can be replaced by a parent test with a statistical guarantee, such as PCM-GAM (Lundborg et al., 2024).

(i) Boundary convention of the theory. All results of the revised manuscript now use the admissible differences of the theory, except the clipped Lahman values kept as a sensitivity analysis. In the synthetic experiments, this changed AtopA_{\mathrm{top}} and F1F_1 by at most 0.02.

(ii) Verification of Assumption 15. Proposition 15A of the revised manuscript verifies Assumption 15 for the empirical score-entropy minimizer on a finite support (see our answer to Q3-1 and Q3-2). On finite supports, the fitted rank transform is the identity with probability tending to one.

(iii) Standardized parent rule and alternatives. The standardized OCS rule is one scalable choice for parent selection, and the order estimated by DISCO can also be combined with a parent test with a statistical guarantee, such as PCM-GAM (Lundborg et al., 2024).

W2-1. Comparison with Vo et al.

The two methods cover different data, and on the same benchmarks DISCO is more accurate on count data, where their method recovers little of the order, while their method is more accurate on categorical data.

(i) Comparison of the data covered by the two methods. DISCO covers count variables in the semiparametric GLM class, whose conditional family is unspecified and may differ across nodes, such as Poisson, negative binomial, and binomial conditionals (Theorem 8). Vo et al. cover categorical variables with finite unordered states, under the condition that conditional randomness, measured for example by conditional entropy, does not decrease along the causal order (Vo et al., 2026, Condition 3).

(ii) Results on the same benchmarks. We ran their released code with their hyperparameters and DISCO on our count benchmark (ER3, d=50d=50, n=5000n=5000) and on their categorical benchmark (Vo et al., 2026, Sections 4 and 5.1), with parents selected by CAM. All entries are means over three seeds, and Time is the ordering time.

DataVo et al. AtopA_{\mathrm{top}}Vo et al. F1F_1Vo et al. Time (min)DISCO AtopA_{\mathrm{top}}DISCO F1F_1DISCO Time (min)
Our benchmark
Poisson0.1320.081197.00.9330.8405.0
NB0.1070.071341.00.9380.8455.0
Binomial0.1420.126165.00.9030.8905.3
Benchmark of Vo et al.
Categorical, d=10d=100.8030.72512.30.7450.6429.3
Categorical, d=20d=200.7760.6509.50.5520.4259.8

Their method refits its diffusion model after each sink removal, which took 165 to 341 minutes per count graph at d=50d=50, so the d=100d=100 benchmarks of the main text were out of reach. We will update these results with further seeds as the runs complete. We added the comparison to Table 6A of the revised manuscript.

W2-2. Adaptation of the ODS baseline

To compare with ODS as originally proposed, all ODS results in the revised manuscript use the exact-conditioning R code that the authors of Park and Raskutti (2018) shared with us, with only the QVF coefficient and the regression link changed for each family. When no conditioning group has enough observations, the original code stops before returning a score, so we set the score to zero. The exact-conditioning code and the regression estimates give F1F_1 of ODS (Oracle) within about 0.02 of each other in every family, so the gap to DISCO does not come from this adaptation.

Q1-1. Relation to the density-ratio literature

We agree that Theorems 5 and 6 relate to this literature. The conditional form of Theorem 5 is the density-ratio model log⁡{f1(y)/f2(y)}=β0+β1y\log\{f_1(y)/f_2(y)\}=\beta_0+\beta_1y of Anderson (1979) [1] and Qin (1998) [2], with their group label in the role of our parent value. The affine non-identifiability of Theorem 6 is the symmetry of log⁡p(x,y)=a(x)+c(y)+β1xy\log p(x,y)=a(x)+c(y)+\beta_1xy behind the equivalence of the prospective and retrospective logistic models of Prentice and Pyke (1979) [3, Section 2]. Our contribution is to characterize the conditional form of Theorem 5 by a zero CCS, which extends the constant-curvature criterion of SCORE. We also show the converse of this symmetry: a nonlinear canonical parameter makes the direction identifiable (Theorem 6), and nonlinear edges make the DAG identifiable (Corollary 9).

References

[1] Anderson, J. A. (1979). Multivariate logistic compounds. Biometrika, 66(1):17–26.

[2] Qin, J. (1998). Inferences for case-control and semiparametric two-sample density ratio models. Biometrika, 85(3):619–630.

[3] Prentice, R. L. and Pyke, R. (1979). Logistic disease incidence models and case-control studies. Biometrika, 66(3):403–411.

Q1-2. Relation to CPCM

In the bivariate case, both identifiability mechanisms are based on the canonical parameter of an exponential family and rest on the same algebraic step, which equates the forward and reverse factorizations of the joint density. The two classes differ in the conditional families they allow and in whether a list of known families must be given, so neither class contains the other. In multivariate DAGs, the algebraic steps differ.

(i) Shared algebraic step. In both classes, the direction is non-identifiable exactly when p(x,y)p(x,y) also factorizes in the reverse direction within the class, and Proposition 5 of Bodík and Chavez-Demoulin (2025) gives sufficient conditions for identifiability, whereas our Theorem 6 gives a necessary and sufficient condition. The same step also appears in the identifiability proofs for additive noise [1], post-nonlinear [2], location-scale [3], and heteroscedastic symmetric noise models (Lin et al., 2025).

(ii) Families covered by each class. Both classes contain one-parameter families such as Poisson, exponential, or Gaussian with fixed variance. Our class leaves the conditional family unspecified, whereas CPCM requires a list of known families. This list lets CPCM cover families with more than one parameter, such as a Gaussian with parent-dependent variance (their Theorem 4 and Consequence 12). We added the statement that neither class contains the other to the related work of the revised manuscript.

(iii) From the bivariate case to DAGs. CPCM extends the bivariate step to DAGs through Theorem 28 of Peters et al. (2014) (their Lemma 11), which requires the bivariate step for every edge given the other variables. Our Theorem 8 instead identifies sinks by a zero CCS on each ancestral remaining set.

References

[1] Hoyer, P. O., Janzing, D., Mooij, J. M., Peters, J., and Schölkopf, B. (2009). Nonlinear causal discovery with additive noise models. In NeurIPS, 21:689–696.

[2] Zhang, K. and Hyvärinen, A. (2009). On the identifiability of the post-nonlinear causal model. In UAI, pages 647–655.

[3] Immer, A., Schultheiss, C., Vogt, J. E., Schölkopf, B., Bühlmann, P., and Marx, A. (2023). On the identifiability and estimation of causal location-scale noise models. In ICML, PMLR 202:14316–14332.

Q2-1. A weaker condition for Theorem 8(i)

As the reviewer reads it, the proof of Theorem 8(i) uses only a per-node condition. We are grateful for this reading, since re-examining the proof with it led us to relax the condition further for count data so that linear edges are allowed (Theorem 9A).

(i) What the proof of Theorem 8(i) uses. The proof of Theorem 8(i) only requires that, for every ancestral set RR and every nonsink jj of GRG_R, the subgraph of GRG_R induced on chGR(j)\mathrm{ch}_{G_R}(j) has a sink c⋆c^\star with Dj2ηc⋆≢0\mathcal{D}_j^2\eta_{c^\star}\not\equiv0. This condition is strictly weaker than edgewise nonlinearity. We state a further count-specific relaxation as Theorem 9A.

(ii) Extension to linear edges in count DAGs. Theorem 9A(i) shows that, for almost every intercept vector, the CCS identifies the sinks of every ancestral remaining set if and only if no node with a single parent has a linear edge into it. This condition is weaker than the per-node condition of the question, since in a v-structure with two linear edges both nonsinks have positive CCS through the log-partition curvature of the child. For the identifiability of GG, Theorem 9A(ii) further shows that only the edges out of sources matter: GG is identifiable if and only if no source has a linear edge into a child with no other parent. The table below gives the largest share of edges that can be linear at once while each condition holds, over 300 random graphs for each setting (Table 8E of the revised manuscript).

GraphddSink characterization holds (9A(i))GG remains identifiable (9A(ii))
ER11066.1%75.8%
5069.0%80.3%
ER31094.2%95.9%
5094.4%97.1%

With the linear and mixed feature maps of Table 8A of the revised manuscript, DISCO recovers most of the order as with nonlinear feature maps, with AtopA_{\mathrm{top}} of at least 0.86 on ER3 graphs with d=50d=50 and n=5000n=5000.

FamilyAtopA_{\mathrm{top}} LinearAtopA_{\mathrm{top}} NonlinearAtopA_{\mathrm{top}} MixedF1F_1 LinearF1F_1 NonlinearF1F_1 Mixed
Poisson0.9050.8990.9100.8210.8050.795
NB0.9130.9160.9150.7770.8030.770
Binomial0.8670.9100.8820.6740.8840.812

Q2-2. Role of Condition R1 at support endpoints

We thank the reviewer for this precise reading, which is correct. Condition R1 is needed only at count-support endpoints. It is used only for linear edges, in the only-if direction of Theorem 6. At interior levels, pY(y)≤1p_Y(y)\le1 and convexity of Ξ~X\widetilde\Xi_X ensure interiority (proof of Theorem 6). At endpoints, R1 is satisfied if the cause has an exponential moment, E[eε∣X∣]<∞\mathbb{E}[e^{\varepsilon|X|}]<\infty for some ε>0\varepsilon>0, covering Poisson, negative binomial, and finite-support causes.

As a counterexample, let pX(x)∝(x+1)−3p_X(x)\propto(x+1)^{-3} on N0\mathbb N_0 and Y∣X=x∼Binomial⁡(M,sigmoid⁡(β0+β1x))Y\mid X=x\sim\operatorname{Binomial}(M,\operatorname{sigmoid}(\beta_0+\beta_1x)), with M≥2M\ge2 and β1>0\beta_1>0. Then Ξ~X=(−∞,β1M]\widetilde\Xi_X=(-\infty,\beta_1M], so the endpoint y=My=M violates R1. Indeed, p(X=x∣Y=M)≍(x+1)−3p(X=x\mid Y=M)\asymp(x+1)^{-3} has infinite variance. Thus, R1 excludes such heavy-tailed endpoint cases rather than standard count causes.

Q2-3. Regularity conditions and what they exclude

The revised main text states the regularity conditions and what they exclude below Table 1 in Section 4, and what the model excludes after Definition 4. Near-deterministic relations remain in the class, since the regularity conditions only require every count level to have positive probability, so they exclude only deterministic relations. Parent-dependent dispersion and zero inflation lie outside the model of Definition 3, where the parents change log⁡p(xj+1∣xpa(j))−log⁡p(xj∣xpa(j))\log p(x_j+1\mid x_{\mathrm{pa}(j)})-\log p(x_j\mid x_{\mathrm{pa}(j)}) only through ηj(xpa(j))\eta_j(x_{\mathrm{pa}(j)}), by the same amount at every level xjx_j. A parent-dependent dispersion changes this difference differently across levels, and zero inflation changes it only between levels 0 and 1.

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

This paper proposes DISCO, a score-matching-based framework for causal discovery from count data without prespecifying conditional distribution families. The framework introduces two curvature-based criteria, the conditional curvature score (CCS) and the off-diagonal curvature score (OCS), to identify causal orderings and recover directed causal edges, respectively. The authors establish identifiability results for nonlinear semiparametric GLM DAGs and develop a discrete diffusion-based algorithm with marginal concrete-score projection. Experiments on synthetic datasets with different count distributions and real-world baseball data demonstrate competitive causal discovery performance, with the method scaling to DAGs containing up to 1,000 nodes.

Critical strengths
  1. The paper addresses an important limitation of existing score-matching-based causal discovery methods, extending their applicability beyond Gaussian additive noise models to a broader class of count distributions.

  2. The proposed CCS and OCS are supported by theoretical identifiability guarantees, providing a principled foundation for recovering causal orderings and complete DAG structures.

  3. The marginal concrete-score projection provides an efficient mechanism for estimating scores across different variable subsets, reducing the computational burden of recursive causal discovery.

Critical weaknesses
  1. The paper proposes CCS-based causal ordering, while Vo et al. (2026) also use discrete score matching, discrete diffusion, and recursive sink identification. However, the paper does not include a direct experimental comparison with this closely related method.

  2. Theorem 11 proves consistency for a variant assuming consistent curvature estimates and raw OCS thresholding, whereas the implementation uses diffusion-based estimation and standardized foldwise selection. The guarantee does not directly cover the implemented algorithm.

  3. The real-data evaluation measures agreement with seven domain-informed relations rather than a ground-truth DAG, some of which may violate the model assumptions. And why 0.000 ordering agreement on raw counts under admissible differences in Table 13 also reports.

Questions for authors

See Weaknesses.

3 comments

We appreciate the reviewer's request for a direct comparison with Vo et al. (2026) and the questions on Theorem 11 and the real-data evaluation. These points led to a head-to-head comparison with this closest method and a stronger real-data evaluation. The revised manuscript also characterizes identifiability for count DAGs with linear edges (Theorem 9A), which requires nonlinearity only on the edges from a source into a child with no other parent. Revised passages of the manuscript are marked in blue.

W1. Comparison with Vo et al.

We now compare the two methods directly, and DISCO recovers most of the order on count data, where the method of Vo et al. (2026) recovers little, while their method is more accurate on their categorical benchmark.

(i) How the two methods differ. The two methods look similar in these three components, but they differ in each of them, most importantly in the model that justifies recursive sink identification. Vo et al. assume that conditional randomness does not decrease along the causal order (Vo et al., 2026, Condition 3), whereas DISCO identifies sinks within the semiparametric GLM class without such an ordering of randomness across nodes (Theorem 8). For discrete score matching, Vo et al. treat the states as unordered categories, whereas DISCO uses concrete scores, the probability ratios of neighboring counts that play the role of the continuous score ∇log⁡p\nabla\log p (Meng et al., 2022). For discrete diffusion, Vo et al. refit a categorical diffusion model after each sink removal, whereas DISCO fits one birth-death diffusion model and projects it to every remaining set.

(ii) Results on the same benchmarks. We ran the code released by Vo et al. with their hyperparameters on our count benchmark (ER3, d=50d=50, n=5000n=5000) and on their categorical benchmark (Vo et al., 2026, Sections 4 and 5.1), with parents selected by CAM on each order. All entries are means over three seeds, and Time is the ordering time.

DataVo et al. AtopA_{\mathrm{top}}Vo et al. F1F_1Vo et al. Time (min)DISCO AtopA_{\mathrm{top}}DISCO F1F_1DISCO Time (min)
Our benchmark
Poisson0.1320.081197.00.9330.8405.0
NB0.1070.071341.00.9380.8455.0
Binomial0.1420.126165.00.9030.8905.3
Benchmark of Vo et al.
Categorical, d=10d=100.8030.72512.30.7450.6429.3
Categorical, d=20d=200.7760.6509.50.5520.4259.8

Because of the refitting after each removal, their method took 165 to 341 minutes per count graph even at d=50d=50, so we did not run it at d=100d=100. We will update these results with further seeds as the runs complete. The comparison is in Table 6A of the revised manuscript.

W2. Gap between Theorem 11 and the implemented algorithm

We agree that Theorem 11 does not cover the standardized OCS rule used in the implementation. The standardized OCS rule is one scalable choice for parent selection, and the order estimated by DISCO can also be combined with a parent test with a statistical guarantee, such as PCM-GAM (Lundborg et al., 2024).

Edit: Although we have not shown Theorem 11 for the neural network used in DISCO, Proposition 15A of the revised manuscript verifies the consistent curvature estimates assumed by Theorem 11 for the empirical score-entropy minimizer on a finite support.

W3-1. Reference relations that may violate the model assumptions

We thank the reviewer for raising the validity of the reference relations. Since a ground-truth DAG is unavailable for the Lahman data, as is common in real-data studies, we initially used seven domain-informed baseball accounting relations as external references for evaluating the estimated order. We acknowledge that these containments, such as H ≤\leq AB, can conflict with the parent-independent support assumption of Definition 3 on raw counts. To broaden the evaluation beyond these containments, we added ten non-containment directions from the MLB analysis of Park and Park (2019b) to the main text. To address the support concern, we examined the data after q4 binning, under which 14 of the 17 reference pairs exhibit complete empirical pairwise support (Appendix H.3). With q4 binning and admissible differences, DISCO attains mean ordering agreements of 0.7620.762 on R7\mathcal R_7 and 0.8040.804 on R17\mathcal R_{17}, compared with 0.3330.333–0.4290.429 and 0.2750.275–0.3920.392 for ODS across family specifications.

W3-2. Cause of the zero ordering agreement on raw counts

The agreement is 0 because DISCO removes AB, H, and BB early as sinks on raw counts, which reverses all seven reference directions: three from AB, three from H, and one from BB. These variables have wide, sparsely occupied supports on raw counts, for example 384 observed levels of AB among 1,400 observations with a median of three observations per level, which makes the neighboring log-ratios and curvatures hard to estimate reliably. To mitigate this estimation difficulty practically, quantile binning reduces this sparsity: after refitting DISCO, under admissible differences the mean AtopR7A_{\mathrm{top}}^{\mathcal R_7} is 0.7620.762 for each of q4, q6, and q8 (Table 13 and Appendix H.3).

Dear Reviewer,

Thank you again for your thoughtful and constructive feedback. We have provided detailed responses to your concerns and incorporated additional experiments and theoretical results into the revised manuscript.

As the discussion period is approaching its end, we would greatly appreciate it if you could share your thoughts on whether our responses have addressed your concerns.

Thank you very much for your time and consideration.

Reviewer_uuLxOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper extends SCORE's sink criterion to count data. SCORE requires a sink's log-density curvature to be constant. This paper instead conditions on the node's own value (the CCS), which removes the curvature coming from the node's own conditional distribution. Combined with an off-diagonal score for parent selection, this gives identifiability for a semiparametric GLM DAG class in which the conditional family does not need to be specified. This class strictly contains SCORE's nonlinear Gaussian ANMs. Estimation uses discrete diffusion together with an amortized projection network for the marginal scores, and scales to 1000 nodes.

Critical strengths
  • The CCS is a simple and natural fix; the NB example makes it very clear why naive finite differences fail.
  • The main results are exact iff characterizations.
  • Not having to choose a conditional family is a real advantage over ODS-type methods.
  • The amortized projection is a nice practical trick, and the experiments are broad and honestly reported.
Critical weaknesses
  1. By Thm 6, an affine η is not identifiable in this class. That includes the standard log-linear Poisson/NB DAG, which ODS can handle. All mechanisms in Table 5 have a nonlinear η, through either non-canonical links or nonlinear feature maps.What is the trade-off? Can you do an experiment on the log-linear case.
  2. The theory and the implementation drift apart in a few places:
    • The rank transform is not affine when the observed support has gaps, so a true sink's CCS need not be zero after the transform.
    • Boundary clipping differs from the admissible differences used in the theory.
    • Thm 11 does not cover the standardized OCS rule that is actually used. As G.5 notes, that rule can never select a parent for a node with only 2–3 predecessors.
  3. The real-data result looks weak. On raw Lahman counts DISCO gets A_top = 0 on R7, while ODS (NB) gets 0.857 (Tables 13–14). The main-text number depends on q4 binning and on boundary handling. This should be discussed in the main text.
  4. OCS is clearly weaker than CAM for parent selection (Table 17), so the 1000-node pipeline trades accuracy for speed.
Questions for authors
  1. How does DISCO do on log-linear Poisson/NB DAGs, or when η is only weakly nonlinear?
  2. With heavy-tailed counts, does the CCS of true sinks stay near zero after the rank transform?
  3. Why does DISCO fail on raw Lahman counts? Is it the n_min filter?
  4. Among the baselines, only DiffAN is score-matching based. SCORE itself appears only in the constant-curvature ablation (Table 2(a)), and Vo et al. (2026), which also targets discrete data. Could you add both as full DAG-recovery baselines (e.g. SCORE with finite differences or on the rank-transformed data)? Additionally, there are a few works based on Score matching on causal discovery are missing, e.g. [1-4]. Theorem 6 excludes affine η. For continuous ANMs, mixtures of linear and nonlinear mechanisms have been handled by first identifying which mechanisms are linear [1]. Could a similar step extend DISCO to the log-linear case?

Refs

[1] Causal discovery with mixed linear and nonlinear additive noise models: A scalable approach CLeaR 2024

[2] Assumption violations in causal discovery and the robustness of score matching. NeurIPS 2023.

[3] A skewness-based criterion for addressing heteroscedastic noise in causal discovery ICLR 2025

[4] Sample complexity bounds for score-matching: Causal discovery and generative modeling. NeurIPS 2023

7 comments

We thank the reviewer for the suggestions of experiments on log-linear count DAGs and of SCORE and Vo et al. (2026) as full baselines. They helped us evaluate DISCO on log-linear count DAGs and compare it with both methods on the same benchmarks. Above all, the question on the log-linear case led us to extend the identifiability result to count DAGs with linear edges (Theorem 9A of the revised manuscript). Revised passages of the manuscript are marked in blue.

W1-1. Trade-off between an unspecified family and an affine canonical parameter

The trade-off is between leaving the family unspecified and allowing an affine canonical parameter. The QVF DAG models of Park and Raskutti (2018) allow an affine canonical parameter but require a known quadratic variance function, Var(Xj∣Xpa(j))=βj0μj+βj1μj2\mathrm{Var}(X_j\mid X_{\mathrm{pa}(j)})=\beta_{j0}\mu_j+\beta_{j1}\mu_j^2, at each node. In particular, for the negative binomial family, the size rr must be known. Our class leaves the family unspecified, and with two variables it needs a nonlinear canonical parameter. With more than two variables, this restriction is substantially weaker, as Theorem 9A of the revised manuscript shows (see our answer to Q4-3).

W1-2 and Q1. Experiments on log-linear and weakly nonlinear DAGs

On log-linear, weakly nonlinear, and mixed DAGs, DISCO recovers the order more accurately than ODS (Oracle), which is given the true conditional families, and than SCORE. This holds in every setting on ER3 graphs and on the very sparse ER1 graphs, where DISCO also has higher F1F_1 except for binomial data with linear feature maps on ER3 graphs. The table compares DISCO with ODS (Oracle) and SCORE on ER3 graphs with d=50d=50 and n=5000n=5000, where the weakly nonlinear map averages the linear and nonlinear maps and mixed DAGs use the linear map on each edge with probability one half.

FamilyFeature mapODS (Oracle) AtopA_{\mathrm{top}}ODS (Oracle) F1F_1SCORE AtopA_{\mathrm{top}}SCORE F1F_1DISCO AtopA_{\mathrm{top}}DISCO F1F_1
PoissonLinear0.8100.6830.3790.2300.9050.821
Weakly nonlinear0.8020.6640.4660.2840.9190.791
Nonlinear0.8220.7050.6490.4400.8990.805
Mixed0.8420.7110.4500.2780.9100.795
NBLinear0.8490.7440.2130.1420.9130.777
Weakly nonlinear0.8440.7610.2910.2020.9170.821
Nonlinear0.8330.7740.4850.3740.9160.803
Mixed0.8370.7480.3010.2080.9150.770
BinomialLinear0.7030.3450.7070.6900.8670.674
Weakly nonlinear0.7910.6430.7310.7090.9050.870
Nonlinear0.8260.7580.7500.7070.9100.884
Mixed0.8010.6410.7460.7220.8820.812

On ER graphs, the gap to ODS is consistent with its sample requirement: the ordering guarantee of ODS needs a sample size of order (log⁡max⁡{n,d})5+Δ(\log\max\{n,d\})^{5+\Delta} (Park and Raskutti, 2018, Theorem 3.9), where the maximum degree Δ\Delta of the moralized graph is between 23 and 37 on the ER3 graphs of the table and between 9 and 12 on the ER1 graphs. SCORE relies on constant curvature at sinks, which fails for count conditionals whose own curvature is not constant (Section 5.1).

However, on SF1 graphs, DISCO has lower AtopA_{\mathrm{top}} than ODS (Oracle) for Poisson and NB data except with nonlinear feature maps, while its F1F_1 remains higher than that of ODS (Oracle) in every setting (Table 8C of the revised manuscript). This agrees with Theorem 9A(i), since 42% of the edges on SF1 graphs enter a node with a single parent, so for Poisson data many linear edges fall in the case that Theorem 9A(i) excludes. These results are in Appendix H.2.7 of the revised manuscript (Tables 8A to 8D).

W2-1 and Q2. Rank transform under support gaps and heavy tails

A true sink can have a nonzero CCS after the rank transform in finite samples, but with finite support this happens with probability tending to zero, and under heavy tails the median CCS of sinks stays below that of the nonsinks at every tested rr.

(i) Support gaps under finite support. Suppose each coordinate has, after the population rank transform, finite support {0,…,Kj}\{0,\ldots,K_j\} and every level vv has positive probability pj(v)p_j(v). A gap occurs only if some level is missed by all nn observations, and by a union bound this has probability at most ∑j∑v{1−pj(v)}n\sum_j\sum_v\{1-p_j(v)\}^n, which tends to zero. Without a gap, the fitted rank transform is the identity on the whole support, so every sink has zero CCS by Theorem 8.

(ii) Sink CCS under heavier tails. We reduced the size rr of the negative binomial setting of Table 5, which makes the counts heavier-tailed, and report the median CCS at the initial sink-removal step.

rrWith rank transform CCS, sinksWith rank transform CCS, nonsinksWith rank transform AtopA_{\mathrm{top}}Without rank transform CCS, sinksWithout rank transform CCS, nonsinksWithout rank transform AtopA_{\mathrm{top}}
60.00040.00320.9360.00040.00320.931
10.00050.00190.8640.00050.00200.863
0.50.00070.00110.7370.00080.00120.771
0.250.00220.00240.7360.00380.00410.665

A nonzero sink CCS does not affect the ordering as long as it stays below the CCS of the nonsinks, since the ordering removes the node with the smallest CCS. The lower AtopA_{\mathrm{top}} at small rr does not come from the rank transform, since it also appears without the rank transform.

W2-2. Boundary clipping versus the admissible differences of the theory

The clipping was only a precaution, and all results of the revised manuscript are recomputed with the admissible differences used in the theory, except the clipped Lahman values kept as a sensitivity analysis. For the synthetic experiments, recomputing the results from the stored models changed every reported AtopA_{\mathrm{top}} and F1F_1 by at most 0.02.

W2-3. Coverage of the standardized OCS rule by Theorem 11

We thank the reviewer for pointing out that Theorem 11 does not cover the standardized OCS rule. The standardized OCS rule is one scalable choice for parent selection, and the order estimated by DISCO can also be combined with a parent test with a statistical guarantee, such as PCM-GAM (Lundborg et al., 2024). The revised rule now selects parents for nodes with two or three candidates, changing the mean F1F_1 of every reported result by at most 0.012. The submitted rule centers each candidate OCS at the median of the candidate scores of its node, which estimates the score of a nonparent, and divides it by their IQR. With two or three candidates, no standardized score can then exceed the threshold. The submitted rule already takes the scale from other nodes when a node has a single candidate, and the revised rule extends this treatment to two or three candidates by also taking the median from the nodes with at least four candidates.

W3. Lahman results on raw counts and their dependence on binning

The raw-count failure is consistent with wide, sparsely occupied supports rather than with the boundary convention, and the revised main text now discusses it together with its dependence on binning.

(i) Why DISCO fails on raw counts. The raw-count failure is consistent with finite-sample estimation difficulties on wide, sparsely occupied supports, related to the support-coverage concerns discussed in W2-1/Q2. For example, At Bats (AB) has 384 observed levels among 1,400 observations, with a median of only three observations per level, which can make neighboring log-ratios and curvatures difficult to estimate reliably. In all three raw runs, DISCO selects AB, hits (H), and walks (BB) early as sinks, reversing all seven reference directions and yielding AtopR7=0A_{\mathrm{top}}^{\mathcal R_7}=0. The reported ODS (NB) agreement of 0.857 comes from our NB extension of Park and Raskutti (2018), which uses a zero-score fallback when no eligible conditioning group is available. In separate raw-count NB diagnostics, this fallback determined 19 of the 48 ordering steps.

(ii) Effect of binning and boundary handling. Quantile binning reduces support sparsity, and after refitting DISCO under admissible differences, the mean AtopR7A_{\mathrm{top}}^{\mathcal R_7} is 0.7620.762 for each of q4, q6, and q8 (Table 13 and Appendix H.3). Clipping changes only the q4 value, to 0.9050.905. The revised manuscript reports the admissible q4 result as the main value and clipping as a sensitivity analysis.

W4. Accuracy of OCS relative to CAM in parent selection

On small graphs, CAM is more accurate than OCS by at most 0.071 in F1F_1 with the order estimated by DISCO, whereas on large graphs only OCS makes parent selection feasible.

(i) Small graphs. On the graphs of Table 17, CAM has higher F1F_1 than OCS in every family with the true order (for example, 0.989 against 0.882 for Poisson data). With the order estimated by DISCO, CAM has higher F1F_1 than OCS for negative binomial (0.823 against 0.754) and binomial data (0.898 against 0.827), and OCS has slightly higher F1F_1 for Poisson data (0.857 against 0.848). CAM took about 14 minutes per graph for parent selection alone, longer than the ordering by DISCO (about 5 minutes).

(ii) Large graphs. We additionally ran CAM on the order estimated by DISCO at d=100d=100 (ER3, Poisson), and it did not finish within 240 minutes in any run, having fitted 84 to 88 of the 99 nodes after the first in the order. OCS completed parent selection up to d=1000d=1000, where the whole pipeline attained F1F_1 of 0.771 in 57 minutes (Table 3).

Q3. Cause of the failure on raw Lahman counts

We thank the reviewer for pointing to the nmin⁡n_{\min} filter as a possible cause. Although the nmin⁡n_{\min} filter can affect CCS estimation, the following diagnostics show that AB remains the first selected sink when the filter is changed. Holding the learned curvatures fixed, we varied nmin⁡n_{\min} and also evaluated regrouping that retains all evaluation observations.

Evaluation ruleAB observations retained (%)First sink
nmin⁡=2n_{\min}=269.4AB
nmin⁡=3n_{\min}=338.0AB
nmin⁡=5n_{\min}=52.3AB
Regrouping (all observations)100.0AB

Thus, the nmin⁡n_{\min} filter does not explain the failure. The result is consistent with the wide, sparsely occupied raw supports, and quantile binning removes the failure (see our answer to W3).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We are grateful to the three reviewers for their close engagement with both the theory and the experiments. Their questions on the theory, the baselines, and the real-data evaluation led to the revisions summarized below, most importantly an extension of our model class that allows linear edges in count DAGs. Revised passages of the manuscript are marked in blue.

  • Allowing linear edges in count DAGs. Theorem 9A shows when count DAGs with both linear and nonlinear edges remain identifiable. Generically, GG is identifiable if and only if no source has a linear edge into a child with no other parent (Section 5.2 and Appendix C.5). Under this condition, about 97% of the edges of ER3 graphs can be linear at once (Table 8E).

  • Experiments on log-linear and weakly nonlinear DAGs. On ER3 and ER1 graphs with linear, weakly nonlinear, and mixed feature maps, DISCO has the highest AtopA_{\mathrm{top}} in every setting, at least 0.82, compared with ODS (Oracle) and SCORE (Tables 8A and 8B). On scale-free graphs, ODS (Oracle) is more accurate in some settings (Tables 8C and 8D).

  • SCORE and Vo et al. as full baselines. On count data, DISCO is more accurate than SCORE in every setting (Tables 6, 6A, and 6B). On count data, it is also more accurate than Vo et al. (2026), with AtopA_{\mathrm{top}} of at least 0.90 against at most 0.15 (Table 6A).

  • Estimation theory. Proposition 15A verifies Assumption 15 for the empirical score-entropy minimizer on a finite support, and Propositions 11A and 15A give stage-wise rates for the CCS and OCS (Appendix E.3). All results now use the admissible finite differences of the theory, except a sensitivity analysis on the MLB data. The revised parent rule also selects parents for nodes with two or three candidates (Appendix G.5).

  • Real-data evaluation. We added ten non-containment directions from Park and Park (2019b) to the main-text evaluation on the MLB data. On the resulting 17 relations, DISCO attains a mean ordering agreement of 0.804, compared with at most 0.392 for ODS across family specifications (Section 7 and Appendix H.3).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.