{"id":"eb3d375c-feef-4449-ab06-de29211ccd5f","arxiv_id":"2505.08343","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"MiCCD learns a causal model of abnormal system data and solves for the lowest-cost intervention that would have prevented the anomaly, outperforming six baselines in experiments.","lead":"This paper presents a decision-making framework that uses what-if reasoning to find the cheapest intervention that restores a broken system to normal. It learns a causal model from abnormal data and then optimizes over possible actions, reporting better results than six existing methods on simulated and real data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Eq. 2 PN constraint is never actually computed: Section 4.2.2's counterfactual estimation yields a single deterministic y*, not a probability, so the claimed minimum-cost x* is undefined.","rationale":"The reader's weakest-assumption statement focuses on the cluster-label identifiability condition, and that concern is legitimate: Lemma 2 requires auxiliary labels that align with distinct noise regimes, and the paper does not verify that GMM clusters satisfy this on real data. However, I judge the more immediately load-bearing issue to be internal and independent of clustering: the objective in Eq. 2 requires a conditional counterfactual probability, yet Section 4.2.2 describes only a deterministic point prediction. The reader noted in the rationale that 'the PN computation is left unspecified,' but did not elevate it to the primary weak point. My stress-test does. Even granting the identifiability of Z, the optimization loop needs a computable estimator for P(Y_do(X=x*) = 0 | X=x, Y=1). The text provides none: no sampling from q(z|x,u), no averaging, no formula mapping the surrogate outputs to the constraint g(x*). Without that, the Lagrangian in Eq. 9 and the SLSQP procedure have no well-defined constraint function. This is a gap in the central claim as stated, not merely an unverified empirical assumption. I therefore retain the reader's REJECT verdict: the empirical direction may be salvageable, but the paper's core optimization claim is under-specified.","tokens_in":16540,"tokens_out":3759,"duration_ms":43303,"concrete_test":"Implement the counterfactual estimator implied by Section 4.2.2 on a 3-node linear SCM X1 -> X2 -> Y with known Gaussian noises. Compute the exact PN(x*) by closed-form or high-accuracy Monte Carlo integration over the true exogenous noise distribution, and compare it with MiCCD's computed PN across 100 candidate intervention vectors. If the procedure only produces a deterministic y*, then the constraint cannot be evaluated as written; if a stochastic estimator is instead used, report its variance and verify convergence to the exact PN values. Additionally, check that the constraint function g(x*) is smooth along a perturbation direction; a step-function g would invalidate the SLSQP-based optimality argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 2 defines the decision problem as minimizing cost subject to PN = P(Y_do(X=x*) = 0 | X=x, Y=1) >= iota. The paper never defines an estimator for this probability. Section 4.2.2's 'Counterfactual Estimation' is deterministic: abduction 'recovers' one noise vector z via the encoder, intervention sets x*, and prediction propagates to 'the counterfactual outcome y*'. That yields a point prediction, not the conditional probability required by the constraint. The variational posterior q(z_j | x_PA_j, x_j, u) from Eq. 8 is never sampled or integrated; there is no Monte Carlo step, no averaging over the posterior noise, and no formula connecting q to the PN value used in the Lagrangian L(x*, lambda) of Eq. 9. SLSQP therefore has no well-defined constraint function g(x*) unless PN is silently replaced by an indicator of a point prediction, in which case g is a step function and the KKT/optimality story is unsupported. This is load-bearing because the headline claim—that MiCCD identifies the minimum-cost intervention vector satisfying the PN constraint—depends entirely on this constraint being computable. The cluster-identifiability gap flagged by the reader compounds the issue, but even granting perfect cluster labels, the optimization objective is not specified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MiCCD, a causal decision-making framework that aims to find a minimum-cost intervention vector x* satisfying a probability-of-necessity (PN) constraint in abnormal states. The framework has two components: a surrogate model that uses GMM cluster labels as auxiliary supervision to recover exogenous noise variables through a causal-structure-aware variational autoencoder, and an optimization stage that uses the surrogate to evaluate counterfactual outcomes and applies SLSQP to minimize intervention cost under the PN constraint. The claimed contributions are identifiability of the counterfactual reasoning and cost-optimal intervention selection; experiments on synthetic data and three real-world datasets compare MiCCD against RCA and cost-based baselines on F1, N-Cost, nDCG@k, and counterfactual r-MSE.","tokens_in":16828,"tokens_out":5407,"duration_ms":52238,"significance":"If the technical claims were valid, MiCCD would be a useful contribution to cost-aware causal decision-making: it couples identifiable noise recovery with continuous optimization over interventions and evaluates on realistic datasets. The paper sensibly builds on external identifiability results (Tahmasebi et al., Khemakhem et al.) rather than attempting a proof from scratch, and the experimental comparison spans multiple baselines and four datasets. However, the central optimization problem is not actually well-defined: the PN constraint in Eq. (2) is never given a computable estimator, and the identifiability argument in Section 4.1.1 contains an unjustified '2d = 2K' claim. Because these issues concern the core of the proposed method rather than presentation, the current manuscript does not support its headline claims.","major_comments":[{"comment":"The paper never defines a computable estimator for the PN constraint P(Y_do(X=x*)=0 | X=x, Y=1) in Eq. (2). The counterfactual estimation procedure in Section 4.2.2 produces a single deterministic outcome y* via abduction, intervention, and prediction; no Monte Carlo averaging or posterior integration over the variational noise q(z_j | x_PA_j, x_j, u) from Eq. (8) is specified, and no formula connects that posterior to the value of g(x*) in Eq. (9). Consequently the constrained optimization problem in Eq. (2) is not well-defined, and the SLSQP/KKT discussion has no constraint function to enforce; this invalidates the central claim that MiCCD identifies a minimum-cost intervention vector satisfying the PN threshold.","section":"4.2.2, Eqs. (2) and (9)"},{"comment":"The step 'at least 2d = 2K weakly separable variables satisfy the identifiable conditions' is not justified. In Lemma 1, K is the number of mixture components and L_w is a property of the mixture; the paper does not define L_w, does not define K in relation to the number of variables d, and gives no argument that each of the d variables contributes exactly two weakly separable distributions. Since the GMM cluster count K is a free hyperparameter that need not equal d, the condition L_w >= 2K cannot be concluded from the stated abnormality assumptions, and the identifiability of the clustering (and hence the auxiliary labels u) is not established.","section":"4.1.1, Lemma 1 and its Implication"},{"comment":"The paper invokes Lemma 2 as the theoretical foundation for noise recovery but never verifies its conditions A1–A3 for the proposed causal graph, decoder class, or cluster-label variable u. In particular, condition A3 requires nk+1 distinct auxiliary values with an invertible matrix L, and the paper does not explain how the GMM cluster labels provide such points or how many clusters are needed. Moreover, even if the latent variables Z are identifiable up to the equivalence in Lemma 2, the paper does not prove that the counterfactual probability PN and the optimal intervention vector x* are invariant under that equivalence; this is a further unstated gap in the claim of identifiable counterfactual reasoning.","section":"4.1.2, Lemma 2"},{"comment":"The empirical evaluation does not specify the intervention cost function C, the PN threshold iota, or the SLSQP settings used to produce the reported N-Cost and F1 results, and the clustering quality is validated only on a synthetic 10-node chain (Figure 7), not on the real datasets. Because the optimization problem is undefined without these components (Major Comment 1), the experimental comparisons cannot be independently reproduced, and the claimed superiority in N-Cost cannot be attributed to the proposed constrained optimization rather than to unspecified choices in the cost function or solver.","section":"5.1, 5.2, and Table 1"}],"minor_comments":[{"comment":"There are typos such as 'anormaly pattern' and 'produce'; please proofread the manuscript carefully.","section":"Section 1"},{"comment":"The notation for the intervention set is inconsistent: Eq. (2) uses x* ∈ R^n while Rule 1 writes X* and X*_R; define the relationship between the vector and the set.","section":"Section 3 and Rule 1"},{"comment":"The text claims an improvement of 'up to 44 percentage points' but the values in Table 1 imply different margins (e.g., 0.94-0.44=0.50 on Lemma-RCA); please recompute or clarify the stated margins.","section":"Section 5.2"},{"comment":"'Thr reconstruction results' and 'nncludes' in Section 5.1 should be corrected; Table D.4's column header formatting (Chain/Sparsity) should align with Table A.3.","section":"Appendix D"},{"comment":"The GMM component is described as providing auxiliary labels, but the number of components K and the initialization/selection procedure are not reported; this is needed to assess whether the assumed one-to-one mapping to anomaly patterns is plausible.","section":"Section 4.2.1"}],"recommendation":"reject","confidential_remarks":"The manuscript is not ready for publication because the central optimization problem is undefined and the identifiability argument contains a mathematical gap. If the authors can supply a precise PN estimator and fix the condition in Lemma 1's implication, a resubmission may be considered. The paper's scope fits the journal, but the novelty claim should be checked against the AUF-MICNS line of work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on 2505.08343. The paper proposes MiCCD, which recovers exogenous noise with a cluster-supervised causal VAE and then uses SLSQP to find a minimum-cost intervention under a Probability of Necessity constraint. The integration is genuinely useful: RCA methods ignore cost, and RL ignores causal structure, so this combination addresses a real gap. The experiments are extensive and the results are consistently favorable across synthetic and three real datasets, including a big win on Lemma-RCA. The authors also correctly lean on identifiability results from Tahmasebi et al. and Khemakhem et al.; the problem setup is honest.\n\nThe central claim does not hold as written. The optimization in Eq. 2 requires PN = P(Y_do(X=x*)=0 | X=x, Y=1) >= iota, but the paper never defines an estimator for this probability. Section 4.2.2 describes deterministic counterfactual estimation: abduct a single z, intervene, predict y*. There is no sampling from the posterior, no averaging, no connection between q(z_j) and the PN value used in the Lagrangian. SLSQP needs a smooth constraint function; if PN is silently replaced by an indicator of a point prediction, the KKT story is unsupported. The stress-test note is right: the objective is undefined, so the claimed minimum-cost x* is not actually specified. That's load-bearing.\n\nThe identifiability argument is also shaky. Lemma 1 requires 2K weakly separable variables; the paper says \"2d = 2K\" where d is the number of observed variables and K is the number of mixture components. Those aren't obviously related. The jump is unexplained, and the cluster labels are never validated against real anomaly patterns. If the GMM clusters don't align with actual noise shifts, the identifiability guarantee fails and counterfactual predictions become unreliable.\n\nThe empirical direction is salvageable—define a proper PN estimator (e.g., Monte Carlo over the posterior), correct the counting, and release code and data. But as written, the main theoretical pillar and the optimization constraint are missing. I'd reject in current form. A serious editor should still send it to peer review because the problem is important and the experiments are extensive; referees might push the authors to fix these gaps. I wouldn't cite it yet.","headline":"The paper's integration is sensible and its experiments are strong, but the central optimization problem is undefined as written—the PN constraint is never actually computed—so the main claim does not hold.","tokens_in":17382,"tokens_out":3232,"would_cite":false,"duration_ms":30938,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that the minimum-cost intervention that resolves an anomaly can be found by recovering hidden noise variables from clustered abnormal data and optimizing a necessity-constrained cost objective.","keywords":["causal decision-making","counterfactual reasoning","identifiability","minimum-cost intervention","root cause analysis","probability of necessity","surrogate model","variational autoencoder"],"falsifier":"Generate synthetic SCMs with $K$ anomaly patterns but fewer than $2K$ weakly separable variables, or merge two distinct noise regimes into one cluster, then fit MiCCD and test whether two observationally equivalent SCMs produce different recovered noise variables and different optimal intervention vectors; divergence would show the identifiability guarantee does not hold in that regime.","tokens_in":16304,"feed_emoji":"🛠️","tokens_out":7864,"duration_ms":71985,"temperature":0.7,"pith_summary":"Decision-making under anomalies is usually split between trial-and-error reinforcement learning, which ignores cost, and root cause analysis, which ignores whether the fix is feasible or affordable. This paper tries to establish a third route: cast the decision as a counterfactual intervention and find the cheapest intervention vector that makes the anomalous target return to normal with high probability. To do that, it must make counterfactual reasoning identifiable from mixed abnormal data, and it does so by clustering abnormal patterns and using the cluster labels to recover the exogenous noise variables of a structural causal model. Given the approximated model, the framework optimizes intervention cost subject to a probability-of-necessity constraint. If the claims hold, operators could get a concrete lowest-cost fix recommendation rather than a ranked list of root causes.","feed_headline":"Counterfactual reasoning finds the cheapest fix for anomalies","feed_subtitle":"MiCCD recovers hidden noise from clustered anomalies, then optimizes intervention cost with a guaranteed fix.","key_machinery":"The load-bearing object is the surrogate causal model: a variational autoencoder whose decoder respects the causal graph, whose encoder inverts the structural equations to recover the exogenous noise $Z$ from each variable, its parents, and the cluster label $u$, and whose training uses an SCM-informed evidence lower bound. The cluster labels come from a Gaussian mixture model over abnormal samples; Lemma 1 guarantees the mixture is identifiable when there are at least $2K$ weakly separable variables, and Lemma 2 guarantees the recovered latent noise is identifiable under an exponential-family conditional prior with enough distinct cluster points. The optimizer is Sequential Least Squares Programming applied to the Lagrangian of the cost function with the probability-of-necessity (PN) constraint, where PN is the counterfactual probability that intervening to $x^*$ flips the anomalous target to normal.","core_discovery":"The central claim is that the minimum-cost intervention vector $x^*$ satisfying $P(Y_{do(X=x^*)} = 0 \\mid X=x, Y=1) \\ge \\iota$ can be identified without online trials, by doing counterfactual reasoning on an estimated structural causal model. The supporting discovery is that the exogenous noise variables $Z$ are identifiable from large amounts of mixed anomaly data when abnormal-pattern cluster labels $u$ are used as supervision: the clusters separate distinct noise-shift regimes, and a causal-graph-structured variational autoencoder trained with those labels recovers the noise variables that drive the anomaly. With the noise recovered, abduction, intervention, and prediction produce counterfactual outcomes for candidate interventions, and SLSQP searches the continuous intervention space for the minimum-cost vector that meets the necessity constraint. On synthetic graphs and on three real datasets, the paper reports that this procedure achieves lower intervention cost and higher F1 and nDCG@k than RCA, LIME, LC, and a cost-aware baseline.","pith_inferences":["Editorial inference: the same abduction-intervention-prediction loop could be applied to personalized treatment selection, where clusters of patient trajectories play the role of abnormal patterns and symptom remission replaces the target threshold.","Editorial inference: the paper's restriction to observed variables is a stated development boundary; a natural extension is to treat unobserved confounders as additional noise components and cluster on their signatures, which would extend the identifiability argument to latent-variable graphs.","Editorial inference: because the identifiability guarantee depends on the number of weakly separable variables, a stress test on real datasets with fewer distinct anomaly regimes could reveal a minimum data complexity below which the cost-optimal intervention is not reliable."],"forward_implications":["Because the intervention search is guided by counterfactual predictions rather than trial-and-error, the framework can recommend fixes in settings where online experimentation is too expensive or unsafe.","The probability-of-necessity constraint acts as a validity filter, so the output is not just a cheap action but one with a quantified chance of actually resolving the anomaly.","The recovered noise variables give per-sample explanations of why the system is abnormal, linking the decision output back to root-cause-style diagnosis.","Experimental results, if taken at face value, suggest the framework keeps its advantage as graphs grow from 5 to 20 nodes and as sparsity and edge weights vary."],"supporting_citations":[{"why":"Defines structural causal models, counterfactual do-semantics, and the probability of necessity that the objective in Eq. 2 is built on.","marker":"Pearl (2009)"},{"why":"Supplies Lemma 1's identifiability condition for finite mixtures, used to justify cluster labels as supervision.","marker":"Tahmasebi et al. (2018)"},{"why":"Supplies Lemma 2's identifiability theory for latent noise in VAEs with auxiliary variables, the basis for recovering $Z$.","marker":"Khemakhem et al. (2020)"},{"why":"Provides the SLSQP algorithm used to solve the constrained minimum-cost intervention optimization.","marker":"Boggs and Tolle (1995)"},{"why":"Provides the Gaussian mixture model used to assign abnormal pattern labels $u$.","marker":"Dempster et al. (1977)"},{"why":"Defines the CausalRCA baseline against which MiCCD is compared in root-cause localization.","marker":"Budhathoki et al. (2022)"},{"why":"Defines the AUF-MICNS baseline that also optimizes minimum-cost changes, the closest comparative method.","marker":"Du et al. (2024)"},{"why":"Supplies the real 5G AIOps causal graph and labeled root-cause data used in evaluation.","marker":"Zhang et al. (2022)"},{"why":"Supplies the air-pollutant dataset with known intervention periods used in the binary intervenability evaluation.","marker":"Zhu et al. (2018)"}],"fun_headline_variants":["Counterfactual model finds lowest-cost anomaly fix","Min-cost causal decisions from counterfactual reasoning","Identifiable counterfactual framework for cheap anomaly fixes","Counterfactual reasoning cuts intervention cost for anomalies","Cost-aware causal fix via identifiable counterfactual logic"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands or falls on the assumption that the GMM cluster labels $u$ correspond one-to-one to the true abnormal noise regimes, with at least $2K$ weakly separable variables; if the clusters do not match the real noise shifts, the recovered noise variables are not identifiable and the counterfactual predictions that drive the optimizer become unreliable.","fun_headline_variants_meta":{"raw":{"variants":["Counterfactual model finds lowest-cost anomaly fix","Min-cost causal decisions from counterfactual reasoning","Identifiable counterfactual framework for cheap anomaly fixes","Counterfactual reasoning cuts intervention cost for anomalies","Cost-aware causal fix via identifiable counterfactual logic"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1379,"prompt_tokens":1004,"completion_tokens":375,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":620,"completion_tokens_details":{"reasoning_tokens":302}},"tokens_in":620,"tokens_out":375,"duration_ms":3628,"temperature":1.0,"reasoning_tokens":302,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:57:30.948796+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate synthetic SCMs with $K$ anomaly patterns but fewer than $2K$ weakly separable variables, or merge two distinct noise regimes into one cluster, then fit MiCCD and test whether two observationally equivalent SCMs produce different recovered noise variables and different optimal intervention vectors; divergence would show the identifiability guarantee does not hold in that regime.","supporting_citations":[{"cited_title":"Variational autoencoders and nonlinear ica: A unifying framework","cited_arxiv_id":null,"evidence_quote":"Supplies Lemma 2's identifiability theory for latent noise in VAEs with auxiliary variables, the basis for recovering $Z$."},{"cited_title":"Sequential quadratic programming","cited_arxiv_id":null,"evidence_quote":"Provides the SLSQP algorithm used to solve the constrained minimum-cost intervention optimization."},{"cited_title":"Causal structure-based root cause analysis of outliers","cited_arxiv_id":null,"evidence_quote":"Defines the CausalRCA baseline against which MiCCD is compared in root-cause localization."},{"cited_title":"Icassp-spgc 2022: Root cause analysis for wireless network fault localization","cited_arxiv_id":null,"evidence_quote":"Supplies the real 5G AIOps causal graph and labeled root-cause data used in evaluation."}],"review_version":1}