{"id":"0dd2c675-c9b5-4531-80eb-0a07d3685649","arxiv_id":"2611.10668","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Cheesefill jointly imputes missing values and learns causal graphs with instantaneous and lagged effects under heteroscedastic noise, using optimal transport and normalizing flows.","lead":"A new method called Cheesefill learns causal structures from time series when values are missing and the noise changes with the parents. It outperforms approaches that treat imputation and graph discovery separately, handling instantaneous effects and heteroscedastic noise.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The practical MMD-relaxed admissibility condition does not directly follow from the theoretical Proposition 3.1; if the MMD penalty does not approximately enforce distributional matching, the optimization may not correspond to the Kantorovich problem, and the theoretical justification for the…","rationale":"The reader correctly identifies the expressivity of the neural family and the MMD relaxation as the weakest assumptions. Our stress-test refines this into a specific gap: the practical objective relaxes the admissibility condition, so the theoretical Proposition 3.1 does not directly justify the algorithm. The proposed concrete test would settle whether this gap matters empirically. Given the paper's strong and transparent empirical results (over 108 synthetic datasets, real-world benchmarks, ablation studies), this concern does not invalidate the central empirical claim; it merely identifies an area where the theory-practice connection could be tightened. Therefore the original ACCEPT verdict remains appropriate, with the suggestion that future work could provide a tighter analysis or a diagnostic for admissibility.","tokens_in":35072,"tokens_out":6153,"duration_ms":59852,"concrete_test":"On the synthetic heteroscedastic 20-node MCAR dataset with 30% missingness used in Figure 3, run a sensitivity analysis: train Cheesefill with λ_map ∈ {0, 0.1, 1, 10}. For each setting, compute the final validation MMD between corrected trajectories and SEM-generated trajectories, as well as instantaneous and lagged F1. If increasing λ_map from 0.1 to 10 changes F1 by more than 5 percentage points, then the MMD relaxation critically influences recovery and the theoretical connection is not operational. If F1 is stable and MMD is below 0.05 (on normalized data), the relaxation is harmless and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's Proposition 3.1 establishes that optimizing over admissible stochastic kernels in an unrestricted family is equivalent to the Kantorovich problem, and that a parametric family yields an upper bound that is tight if it contains an ε-optimal admissible kernel. However, the practical algorithm replaces the exact admissibility constraint (Definition 2.1) with an MMD penalty between corrected trajectories Y and SEM-generated samples Y_SEM. This penalty does not guarantee that T_ψ ∈ F(P_\\tilde{X}, Q_Θ); it only encourages distributional matching in the MMD sense, which is not equivalent to the exact marginal constraint. Consequently, the theoretical guarantee in Eq. (5) does not directly apply to the optimized objective. The paper acknowledges this relaxation as a limitation but does not quantify its impact. If the MMD penalty is weak or the optimization fails to reduce it sufficiently, the corrected trajectories may remain inconsistent with the SEM, and the graph recovery could be driven by the imputer rather than the causal mechanisms. This concern is load-bearing because the central claim that Cheesefill recovers causal structure relies on the correction map being compatible with the SEM; the theoretical framing claims such compatibility; yet the practical enforcement is indirect and unvalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper introduces Cheesefill, a framework for causal structure learning from incomplete multivariate time series (MTS) that jointly handles missing data, instantaneous and time-lagged causal links, and heteroscedastic noise. The approach couples a parametric imputer, a stochastic correction map (modeled as a neural network), a temporal graph neural network for the mean functions, and conditional normalizing flows for the heteroscedastic residual densities. The key theoretical result (Proposition 3.1) shows that optimizing over admissible stochastic correction maps is equivalent to a conditional kernel reformulation of the Kantorovich optimal transport problem, and that a neural parameterization yields an upper bound that is tight if the family contains an ε-optimal admissible kernel. The practical objective relaxes the exact admissibility constraint with an MMD penalty. Experiments on synthetic data (MCAR, MAR, MNAR, varying missing rates and noise levels) and two real-world benchmarks (NetSim, Wind Tunnel) demonstrate consistent improvements over several baselines.","tokens_in":35462,"tokens_out":2331,"duration_ms":22584,"significance":"If the reported results hold, Cheesefill represents a meaningful advance because it is, to my knowledge, the first method to simultaneously address three realistic challenges—missing observations, instantaneous effects, and heteroscedastic noise—in a single framework. The optimal transport perspective is novel in this application and provides a clean conceptual link between imputation and causal discovery. The ablation study convincingly shows that both the stochastic correction map and the conditional normalizing flow are necessary for the strong performance. The experimental evaluation is extensive (>108 synthetic datasets, two real benchmarks) and the code is provided, supporting reproducibility.","major_comments":[{"comment":"The central theoretical justification (Proposition 3.1) for the stochastic correction map relies on the exact admissibility condition (Definition 2.1). The practical algorithm, however, replaces this condition with an MMD penalty (Section 4, last paragraph and Eq. (10)). The authors acknowledge this relaxation as a limitation (Section 6, Appendix B), but they do not analyze how the MMD penalty relates to the theoretical optimal transport bound. Concretely, the bound in Eq. (5) holds only for admissible kernels; when the MMD penalty is minimized but not zero, the optimized ψ may not lie in Ψ_adm, and Eq. (5) no longer bounds the Kantorovich cost. This gap is load-bearing because the paper motivates the correction map by Proposition 3.1, yet the algorithm's objective is a relaxation whose proximity to the Kantorovich problem is unquantified. I request that the authors either (a) provide an analysis (e.g., a bound on the gap in terms of the MMD error) or (b) soften the claim that the theoretical result directly justifies the practical objective, and discuss under what conditions (e.g., when λ_map is large enough) the relaxation is expected to be approximately valid.","section":"§3.2 (Eq. (6) and §4, Eq. (10) and §6 (Limitations)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The correction map T_ψ is parameterized as a diagonal Gaussian over missing entries only, with fixed identity covariance. The authors justify this choice by noting that the CNF already models parent-dependent residual scale. However, the expressiveness of this family is crucial for the ε-optimal admissibility condition in Proposition 3.1. The paper does not discuss whether this restricted family can approximate an admissible kernel to within a small ε for the problems considered. I recommend adding a remark about the expressiveness limitation and, if possible, a sensitivity analysis (e.g., varying the covariance or allowing non-diagonal structure) to assess the practical impact.","section":"§4 (Correction map parametrization)","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"The conclusion contains a typographical artifact: 'noise regimes.]' has a stray closing bracket after the period. Please correct.","section":"Abstract and §6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The related-work paragraph lists 'neural approaches' for Granger causality; consider citing a more recent survey or standard reference for completeness.","section":"§1 (Introduction, first paragraph)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The notation in Figure 1 uses 'G→' twice and the arrow label 'G→' could be clearer; consider using distinct labels for the graph (G*) and the TGNN/CNF parameters (Θ).","section":"Figure 1","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"After Eq. (5), the text says 'The inequality is tight whenever the parametric family contains an optimal admissible kernel.' This is correct, but for precision, consider noting that the family must also be able to represent that kernel exactly (not just in the limit).","section":"§3.2 (Eq. (5))","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The ablation table (Table 1) reports F1, precision, recall, and ratio. What does 'Ratio' refer to? Please define in the caption or text.","section":"§5.1 (Table 1)","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The MNAR experiments use 'a small observed subset to model the missingness process'. It would be helpful to specify exactly which 10% of variables are fully observed and how this choice was made (e.g., fixed random seed).","section":"Appendix G (MNAR Experiments)","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"The paper tackles a genuinely hard open problem: causal structure learning from incomplete multivariate time series when you have both instantaneous and lagged links plus heteroscedastic noise.  The combination of optimal transport, a stochastic correction map, and conditional normalizing flows is novel and well motivated.  The experiments are extensive—over 100 synthetic settings plus two real benchmarks—and the results are consistently strong across MCAR, MAR, and even MNAR.  The ablation studies confirm that both the correction map and the normalizing flow matter.\n\nThe main soft spot is a gap between theory and practice.  Proposition 3.1 gives a clean equivalence: optimizing over admissible stochastic kernels is exactly the Kantorovich problem.  But the practical algorithm replaces the exact admissibility constraint with an MMD penalty.  That is a relaxation, and the theoretical upper bound (Eq. 5) no longer applies directly to the optimized objective.  The authors acknowledge this in the Limitations section, but they do not quantify the impact—no sensitivity analysis on the MMD weight, no check of how well the penalty actually enforces the constraint.  For a paper that leans heavily on the OT formalism, this is a real and load-bearing gap.\n\nThat said, the empirical evidence is strong enough that the method works despite the relaxation.  The paper is a genuine step forward for practitioners.  What it delivers is a well-engineered neural pipeline that outperforms existing methods; the theory is a useful framing but not a provable guarantee in practice.\n\nWho is it for: researchers working on causal discovery from time series, especially those dealing with real-world missingness.  The paper deserves a serious peer review.  I would recommend accept with revision, asking the authors to either (a) provide an experiment that varies the MMD weight and shows how the recovered graph degrades, or (b) add a formal argument that the MMD relaxation still leads to a consistent estimator under some conditions.  Without that, readers may overinterpret the theoretical result.","headline":"Solid method for a hard problem, but the theoretical framing overpromises relative to the practical relaxation.","tokens_in":35860,"tokens_out":1178,"duration_ms":13211},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Cheesefill jointly recovers causal graphs, imputes missing values, and models heteroscedastic noise from incomplete time series.","keywords":["causal structure learning","multivariate time series","missing data","heteroscedastic noise","optimal transport","stochastic correction map","conditional normalizing flows","Bayesian structure learning"],"falsifier":"If one can construct a dataset with known ground-truth graph, heteroscedastic noise, and 40% MCAR missingness where Cheesefill's F1 score drops below that of a simple impute-then-Fantom pipeline, the claim of consistent outperformance would be challenged.","tokens_in":36625,"feed_emoji":"🔗","tokens_out":4045,"duration_ms":36879,"temperature":0.7,"pith_summary":"Cheesefill is a framework for causal structure learning from multivariate time series that simultaneously handles three real-world challenges: missing values, instantaneous (contemporaneous) and time-lagged causal links, and heteroscedastic noise (noise variance that depends on the cause values). The key idea is to frame causal discovery under missingness as an optimal transport problem between the distribution of naively imputed trajectories and the distribution induced by a structural equation model. A learned stochastic correction map refines imputations to make them consistent with the causal mechanisms, and a conditional normalizing flow models the parent-dependent noise. Theoretically, optimizing over such correction maps is equivalent to a conditional kernel formulation of the Kantorovich problem. Experiments show that Cheesefill consistently outperforms impute-then-discover pipelines and existing methods designed for incomplete MTS across various missingness regimes, noise types, and graph sizes.","feed_headline":"Cheesefill recovers causal graphs from 50% missing data","feed_subtitle":"An optimal transport framework learns imputation and heteroscedastic causal mechanisms jointly, beating all baselines.","key_machinery":"The stochastic correction map $T_\\psi(dY|\\tilde{X})$ refines pseudo-complete trajectories $\\tilde{X}$ from a parametric imputer into refined trajectories $Y$ that are consistent with the SEM. Proposition 3.1 proves that optimizing over admissible correction maps is equivalent to the Kantorovich optimal transport problem between the pseudo-complete and SEM-induced distributions, with the mechanism cost (negative log-likelihood under the SEM) as transport cost. The correction map is parameterized as a diagonal Gaussian over missing entries only, conditioned on the pseudo-complete trajectory. The SEM uses a temporal graph neural network for mean functions $f_i$ and conditional normalizing flows for residual densities $\\rho_i$, capturing heteroscedasticity. Bayesian structure learning averages over graph samples from a variational posterior, with entropy regularization.","core_discovery":"This paper claims that it is possible to recover both instantaneous and time-lagged causal links from incomplete multivariate time series in the presence of heteroscedastic noise, by jointly learning an imputation strategy and a causal model in an optimal-transport framework. The central contribution is showing that a stochastic correction map that refines naive imputations can be interpreted as a conditional kernel reformulation of the Kantorovich problem, making the joint optimization theoretically principled. The method, Cheesefill, parameterizes the correction map and the causal mechanisms (mean functions and heteroscedastic noise distributions) with neural networks, and uses Bayesian structure learning over graphs to mitigate brittleness under missingness. Extensive empirical results demonstrate robust performance across MCAR, MAR, and MNAR missingness, with up to 50% missing entries.","pith_inferences":["If the optimal transport framing is as fundamental as claimed, then any future method that replaces the neural correction map with a different coupling (e.g., a learned deterministic transform or a GAN-style discriminator) should be measurable against the same Kantorovich equivalence.","The paper's reliance on an MMD penalty to enforce admissibility of the correction map is a practical relaxation that may not guarantee distributional matching in high dimensions; this suggests a need for tighter surrogates, possibly based on Sinkhorn divergences or adversarial objectives.","The Bayesian structure learning component could be extended to non-stationary settings by allowing the variational posterior to vary over time or regimes, which the authors mention as a limitation."],"forward_implications":["Cheesefill enables reliable causal discovery from real-world time series where missing data and heteroscedastic noise are common, such as fMRI, EEG, and financial data.","The optimal transport perspective suggests that other imputation-sensitive tasks (e.g., counterfactual inference, treatment effect estimation) could benefit from a similar joint correction-and-discovery approach.","The method's robustness to MCAR, MAR, and MNAR missingness (up to 50%) opens the door for causal analysis in datasets previously considered too incomplete.","By recovering both instantaneous and lagged links, Cheesefill provides a more complete picture of causal dynamics than methods restricted to lagged relationships.","The ablation study confirms that both the stochastic correction map and the heteroscedastic noise model are essential, implying that simpler pipelines will likely fail in the combined challenging setting."],"supporting_citations":[{"why":"PCMCI+ is the main baseline for constraint-based causal discovery with contemporaneous links on complete data.","marker":"[36]"},{"why":"Fantom provides the closest complete-data method for heteroscedastic noise modeling with flows, which Cheesefill extends to the missing-data setting.","marker":"[32]"},{"why":"Rhino supplies the history-dependent noise SEM and Bayesian structure learning strategy that Cheesefill adapts.","marker":"[14]"},{"why":"CUTS is the primary baseline for causal discovery from incomplete MTS, though it is limited to lagged relationships and homoscedastic noise.","marker":"[8]"},{"why":"Conditional spline flows provide the backbone for modeling heteroscedastic residual densities within Cheesefill's SEM.","marker":"[10]"},{"why":"The differentiable acyclicity penalty (DAGs with no tears) is used to enforce the instantaneous DAG constraint.","marker":"[46]"}],"fun_headline_variants":["Joint learning beats missing data for causal graphs","Cheesefill: causal discovery from incomplete time series","Optimal transport fixes missing data for causal inference","Cheesefill handles missing values and heteroscedastic noise","Causal structure learning with 50% missing data"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The neural network family used for the stochastic correction map must be expressive enough to contain a kernel that perfectly matches the pseudo-complete distribution to the SEM-induced distribution; otherwise the optimal transport bound is not tight and the method may not achieve the theoretical optimum.","fun_headline_variants_meta":{"raw":{"variants":["Joint learning beats missing data for causal graphs","Cheesefill: causal discovery from incomplete time series","Optimal transport fixes missing data for causal inference","Cheesefill handles missing values and heteroscedastic noise","Causal structure learning with 50% missing data"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":5.9e-05,"raw_usage":{"cost":5.9248644e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":5.9248644e-05,"upstream_inference_prompt_cost":2.9819244e-05,"upstream_inference_completions_cost":2.94294e-05},"total_tokens":1370,"prompt_tokens":985,"completion_tokens":385,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":307}},"tokens_in":985,"tokens_out":385,"cost_source":"openrouter_usage","duration_ms":3922,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":307,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-11-10T21:44:32.479771+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"If one can construct a dataset with known ground-truth graph, heteroscedastic noise, and 40% MCAR missingness where Cheesefill's F1 score drops below that of a simple impute-then-Fantom pipeline, the claim of consistent outperformance would be challenged.","supporting_citations":[{"cited_title":"Discovering contemporaneous and lagged causal relations in autocorrelated nonlinear time series datasets","cited_arxiv_id":null,"evidence_quote":"PCMCI+ is the main baseline for constraint-based causal discovery with contemporaneous links on complete data."},{"cited_title":"Flow based approach for dynamic temporal causal models with non-gaussian or heteroscedastic noises","cited_arxiv_id":null,"evidence_quote":"Fantom provides the closest complete-data method for heteroscedastic noise modeling with flows, which Cheesefill extends to the missing-data setting."},{"cited_title":"Rhino: Deep Causal Temporal Relationship Learning With History-dependent Noise","cited_arxiv_id":"2210.14706","evidence_quote":"Rhino supplies the history-dependent noise SEM and Bayesian structure learning strategy that Cheesefill adapts."},{"cited_title":"CUTS: Neural Causal Discovery from Irregular Time-Series Data","cited_arxiv_id":"2302.07458","evidence_quote":"CUTS is the primary baseline for causal discovery from incomplete MTS, though it is limited to lagged relationships and homoscedastic noise."},{"cited_title":"Neural spline flows","cited_arxiv_id":null,"evidence_quote":"Conditional spline flows provide the backbone for modeling heteroscedastic residual densities within Cheesefill's SEM."},{"cited_title":"Dags with no tears: Continuous optimization for structure learning.Advances in neural information processing systems, 31, 2018","cited_arxiv_id":null,"evidence_quote":"The differentiable acyclicity penalty (DAGs with no tears) is used to enforce the instantaneous DAG constraint."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}