acceptodds
Under review as a conference paper at ICLR 2027

Beyond Conditional Independence: Root Cause Analysis with Deep Causal Models

Abstract

Root cause analysis (RCA) is a critical problem in many real-world scenarios. RCA enables the identification of faulty or failing mechanisms in a system by comparing anomalous observations with corresponding reference (i.e., regular) observations. However, existing approaches rely either on heuristic methods or on conditional independence tests with a strong unconfoundedness assumption, and thus fail to exploit other complicated distributional constraints in the presence of latent variables. To relax these assumptions, we model the underlying system as a causal model and the anomalous system as a change in the structural functions of the same causal model. Specifically, to handle unobserved confounders, we establish an implicit connection between distributional constraint testing and root cause analysis. To adapt our approach to data generated from arbitrary causal models, we employ the deep causal model (DCM) framework, in which we design the causal model using neural networks. Finally, we illustrate how our method, RCA-DCM, can utilize different levels of partial graphical knowledge to perform RCA. We evaluate RCA-DCM against state-of-the-art baselines on simulated datasets, a physics-based causal chamber, and two microservice applications. RCA-DCM improves top-1 accuracy over the strongest baseline on both Sock Shop (0.880 vs. 0.752) and Online Boutique (0.776 vs. 0.712), and when the true root cause in the causal chamber is unobserved and acts as a latent confounder, it recovers the exact root-cause set more often than any competing method (PRR 0.846 vs. 0.731).

Primary area:
causal reasoning
Keywords:
causal inference, root cause analysis, deep causal models, unobserved confounding, mechanism shift

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_8DmMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies RCA from a normal and anomalous dataset when the system can contain unobserved confounders. The anomaly is a soft intervention (Ikram et al, 2022) encoded by an indicator F with edges to the RC R* . The authors define a RC solution set as SolRc(G): the set R such that the augmented graph can realise (Pn, Pa). They argue that inequality constraints such as Pearl's instrumental inequality can shrink it beyond what CI test allow. they do not derive these constraints explicitly. Instead for each variable Y they train two deep causal models (causal normalizing flows) with functions fy frozen across regimes and all other mechanisms free. If the population loss l* is positive, Y must be a root cause (Theorem 4.13) and variables are ranked by the final loss. Experiments cover: two small synthetic settings (6 observed + 4 latent variables; 5 or 10 variables w/o latents), a graph-perturbation study, the light tunnel causal chamber (38 variables), a sock shop and online boutique from RCAeval.

Critical strengths

The question matters and the core idea is simple. The framing (two regimes plus an indicator F) is standard since Ikram et al. (2022). What is new is the test: ask whether any SCM pair with fy frozen can explain the shift. That one test implicitly covers equality and inequality constraints, so they don't have to be enumerated.

This is a way to go beyond CI-based RCA under hidden confounding. The bow/instrument example (Ex. 4.1, Fig. 1) explains clearly why CI tests are not enough, and the problem statement (l.155-161) is precise.

Against the listed baselines, the method is consistently strong where confounding is present. For instance the method is first in top-1 accuracy on both microservice benchmarks. If the concerns below are resolved, this could be a useful practical tool.

Critical weaknesses

Gap between the guarantees and what is actually run, with assumptions that are not stated at all

Lemma 4.12, Thm 4.13 and Prop 4.14 are all about the population loss ℓ∗\ell^* of Eq. (2). The algorithm instead computes a score from flows of fixed capacity, trained by SGD on finite samples. Going from one to the other requires (&) the DCM is "representative enough" (l.279-281), (2) non-convex training actually reaches the infimum and (3) the empirical loss concentrates around the population loss.

Neither 1 or 2 appears in the assumption list (App. C). Only 3 is treated, in App. E, and Prop E.1 simply assumes "model training is perfect". Optimisation error is not a uniform offset across candidates, since each candidate freezes a different mechanism. It can therefore reorder the ranking!, and the ranking is what gets evaluated. Small anomalous windows and rare events make this worse: Table 2 shows PRR falling from 0.63 to 0.27 when Na=100N_a = 100. This has to be clearer in the revised version.

The evaluation does not support the claims

Scale: Exp 1 has p=6p = 6 observed variables (+4 latents), and Exp 2 has n∈{5,10}n \in \{5, 10\}, with only two points per axis. In Exp 2, PRR already falls from 98% to 54% between n=5n = 5 and n=10n = 10. Runtime is about 41 s per dataset at 12 variables, with one DCM per candidate (Table 4). Scalability is therefore open, and that is a fair question in 2026 I believe. The Fig. 4 caption doesn't say that (a) and (b) are synthetic, or give pp, mm and NN.

Baselines: None of the baselines handles latent confounding, yet that is precisely the paper's claim. Jaber et al. (2020), which the paper cites, learns unknown soft-intervention targets in the presence of latents. Yang et al. (2024), whose generator is used for the synthetic data, identify intervention targets in nonlinear SCMs with latent confounders. Neither is compared or discussed in the related work, so the sentence at l.126-127 ("none explicitly models nonlinear mechanisms with latent confounders") is not accurate and particularly unprecised

Microservices: Unless I misread the construction, take a single latent confounding all non-sink services, together with latent invariance. Freezing fYf_Y for a non-sink YY then only forces Pn(Y)=Pa(Y)P_n(Y) = P_a(Y), because the latent is rich enough to fit any joint distribution with that marginal. The test then reduces to a per-service comparison of YY's own distribution across regimes. That is the marginal-shift heuristic criticised in Sec. 1 (l.073-076, and it would explain why the results are close to NSigma. The conclusion at l.475-476, that the advantage "does not depend on access to an accurate structure", should be revisited in this light

Where the structure is known: In the Causal Chamber with the root cause observed and no confounding, RCA-DCM is behind RCG and RCD (88 vs. 100 / 90). The least favourable results (sample size in Table 2, runtime in Table 4) appear only in the appendix, this merit a more prominent discussion or acknowledgment

Minor remarks

I found when reading painful presentation problems which explain my low score on presentation. I feel that the authors put some effort in the writing but this is still not enough and requires substantial modifications. For instance:

  • l.133-146. Def 3.1 introduces an SCM without a graph. Def 3.2 says it "induces" a DAG over V∪UV\cup U that is never named, then uses G=(V,E)G=(V,E) for the ADMG. "ADMG" is never spelled out, and the latent projection is not cited (Verma & Pearl 1990; Richardson et al. 2023). Def 3.1 also gives one UiU_i per variable with a product distribution PP, which does not match shared confounders. V\mathbf V is written three different ways.
  • l.147. "mDAG" is used without a definition...
  • Notation clashes. M\mathcal{M} is both an SCM (Def 3.1) and a set of distributions (Def 4.2), so "P∈M(⋅)P\in\mathcal{M}(\cdot)" reads as a distribution belonging to an SCM. G\mathcal{G} denotes a neural network (Def 4.11, Alg. 2) while GG is the graph.
  • Def 4.2 is not a sentence, and "classical unrestricted semantics" is undefined. Defs 4.4 and 4.5 come from Evans / Ansanelli et al. but are not cited, whereas Lemma 4.6 is. "Latent facet" is not defined for ADMGs.
  • Too many definitions and results. Cor 4.16 restates Lemma 4.15. C.4, C.6 and D.3 say the same thing, and C.5 is already part of C.2. Most results (4.8, 4.12-4.18) follow in one line from the definitions plus Lemma 4.6. I would keep the main theorem in the text and move the rest to the appendix.
  • Proof sketches. The main text gives no sketch or intuition for 4.15-4.18. The intuition for 4.13 (l.298-300) is hard to parse and writes ℓ∗(G∗,Y)\ell^*(G^*,Y) instead of ℓ∗(G∗,V∖{Y})\ell^*(G^*,\mathbf V\setminus\{Y\}).
  • l.167. Point "Graph 1" to Fig. 2(a).
  • Fig. 2 leaves the bidirected arcs unlabelled, while Fig. 1(a) labels UU. Please make them consistent.
  • Ex. 4.1. The two-node bow works for the IV argument. RCA, though, is usually about a shift that propagates from an upstream cause to an observed symptom (e.g. X→Z→YX\to Z\to Y). A small chain with a confounder would be more convincing here.
  • l.287-291. "absence of the edge F→RF\to R" and "R must be a root cause" should use YY. On l.294, "then ⇒" is redundant.
  • Cross-references. l.408, l.459 (the parenthesis is never closed) and l.474 refer to "the Appendix" without a section. Exp 3 has no table or figure. Each experiment should point to its result.
  • Reproducibility statement (l.488-494). The assumptions are in App. C, not B. No completeness result exists. Complexity is in G.6, not G.3.
  • Naming. "Exp 6" is used twice (l.462, l.1142). The method is called RCA-DCM, DCM-RCA and DCM.
  • References. Ikram 2022a/b, Ikram 2025a/b and Li 2022a/b are duplicates. Okati et al. 2024 and Orchard et al. 2025 are the same paper (arXiv and NeurIPS versions) but are discussed as two works (l.50, l.757).
  • Exp 1. The claim of more robustness than RCD is not supported: both drop 15 points.
  • l.792-794. Assumption C.6 is garbled ("minimal element.necessarily R∗R^*").
  • Conclusion (l.481-484). It is four lines long and does not mention the microservice results. It refers to an "assumption on partial order" that the main text never states. Please move the limitations (graph access, runtime, sample size, identifiability) from App. A.1 into the main text.
Questions for authors

The answers could change my assessment and I will be happy to engage into the discussion. Here are my questions :

  1. Scoring: Alg. 1 scores candidates with the training loss LL. the flows are apparently trained by maximum likelihood and candidates are ranked by a Wasserstein distance between samples and data. Which is it? Is the Wasserstein distance computed on the joint over V\mathbf{V}, or per variable? If per variable, joint constraints such as the IV one cannot affect the score at all.

  2. Training: How is maximum likelihood done with shared latent noise in a causal normalizing flow? The appendix also gives three descriptions of the training: two DCMs per candidate, each DCM "fit solely to its own sample" and one normal DCM plus one per candidate. Which is correct, and how exactly is fYf_Y shared?

  3. Microservices: What is the "service under test" that all services point to? If it is the injected service, that leaks the label. Can you confirm or refute my reading that, for non-sink nodes, the test reduces to a per-service marginal comparison? Please also report RCA-DCM with the true call graph, as CIRCA and RCG receive.

  4. Inconsistent numbers. The abstract (l.029) reports PRR 0.846, while Fig. 4(c) and l.442 report 87%. Also, l.443 says "−2 points" but the figure shows 0.88 → 0.87. In Exp 3, the true-graph control gives 88% and 84% in two runs on the same data. Are the main results from a single seed?

  5. Causal Chamber: Why do NSigma and BARO reach only 6% with a single strong target observed? Is the intervened variable sometimes dropped as constant ?

6 comments
AuthorsOpenReview ↗

Rebuttal Part 1/3

We thank the reviewer for the careful review. Below, we address their concerns.


Scale.

"... In Exp 2, PRR falls from 98% to 54% ... Scalability is therefore open ..."

We would like to point out that the drop in Exp. 2 is not a scalability issue: we increase the number of variables with a fixed sample size of 5k, which leads to positivity violations and sample uncertainty. For large graphs, if we increase the sample size as well, the performance improves. Based on the reviewer's suggestion, we performed an experiment with a random DAG (20 observed variables, 91 edges) having 10 root causes, with $N=$5k-50k samples per regime. We report

T@10=1D∑d=1D∣Top10(d)∩Rd∗∣10T@10=\frac{1}{D}\sum_{d=1}^{D}\frac{|Top_{10}(d)\cap R^*_d|}{10}, where D=30D=30

is the number of datasets, Rd⋆R_{d}^{\star} the true root-cause set of dataset dd, and Top10(d)Top_{10}(d) the method's top 10 nodes for that dataset:

Method5k20k50k
RCA-DCM0.7770.8230.887
RCD0.8370.47timeout
RCG0.6400.6830.853
CIRCA0.6600.6400.763
NSigma0.6800.6830.740
BARO0.5930.6230.740

The conditional independence tests in RCD time out for 20 variables with 50k samples. RCA-DCM performs well when enough samples are provided.


Baselines with latent confounding.

"... Jaber et al. (2020) ... Yang et al. (2024) ... Neither is compared ..."

Based on the reviewer's suggestion, we compared RCA-DCM with Ψ\Psi-FCI (Jaber et al., 2020) and LIT (Yang et al., 2024):

MethodChamber obs. (PRR)Chamber hidden (PRR)Sock Shop (T@1)Online Boutique (T@1)
RCA-DCM0.8850.8850.8880.776
Ψ\Psi-FCI0.9230.8460.6640.712
LIT, linear0.1730.6920.2640.416
LIT, nonlinear0.5000.1350.1840.200

Ψ\Psi-FCI is the strongest of the two. It is ahead in the observed chamber condition, where RCA-DCM misses 6 cases because a child's shift falls outside the input range of the flow, and behind everywhere else. LIT performs worse, as it requires more than the two environments available in RCA. We will discuss both methods in the related work and rewrite l.126-127.


Gap between the guarantees and what is run.

"... requires (1) the DCM is 'representative enough', (2) training reaches the infimum and (3) the empirical loss concentrates ..."

(1) For the DCM being representative enough, we follow existing work on deep causal models [1,2,3], which shows that neural networks can represent the SCMs compatible with a given causal graph. (2) That non-convex training reaches the infimum is a standard assumption when training neural networks with SGD. We will state both assumptions in App. C. (3) For the concentration of the empirical loss, Prop. E.1 gives the condition under which every root cause is ranked above every non-root under sampling error.

"... Optimisation error is not a uniform offset across candidates ... Small anomalous windows ... make this worse ..."

Prop. E.1 easily extends to approximation and optimization error, including errors that differ across candidates: the ranking is correct if min⁡Y∈R⋆ℓ⋆(Y)>2ϵstat+max⁡Z∉R⋆[a(Z)+o(Z)]\min_{Y\in R^{\star}}\ell^{\star}(Y)>2\epsilon_{stat}+\max_{Z\notin R^{\star}}[a(Z)+o(Z)], where a(Z),o(Z)≥0a(Z),o(Z)\ge0 are the approximation and optimization errors of a non-root ZZ. Small anomalous windows increase ϵstat\epsilon_{stat}, which is consistent with the drop in Table 2. We give the detailed argument in our response to Reviewer PsX8 (W1/Q2).


Microservices.

"... take a single latent confounding all non-sink services, together with latent invariance."

We would like to humbly point out that this is a misunderstanding. We perform ∣V∣|V| tests for ∣V∣|V| variables independently. Whenever we test a variable YY, we assume it is a sink node: a confounded sink, in which every other service points to YY and a single shared latent confounds all of them. Thus, the tested variable is never a non-sink, and the marginal test does not happen for it.


Where the structure is known.

Although RCA-DCM outperforms the baselines in confounded setups, we do not claim this for all possible setups. We agree with the reviewer that when the root cause is observed and there is no confounding, some graph-based baselines perform better. We will explicitly mention this in the main paper, together with the limitations on sample size and runtime.

AuthorsOpenReview ↗

Rebuttal Part 2/3

Minor remarks.

We thank the reviewer for pointing out these issues. We will fix them in the camera-ready version. On the remaining points:

"Def 3.1 introduces an SCM without a graph."

To the best of our knowledge, the structural model does not require a graph. An SCM is qualitatively represented with a causal graph (Def. 3.2).

"Most results ... follow in one line from the definitions plus Lemma 4.6 ..."

We respectfully disagree. Lemma 4.12, Theorem 4.13 and Proposition 4.14 are about how the DCM searches for SCMs for root-cause detection. Lemma 4.15, Corollary 4.16 and Lemmas 4.17-4.18 are about how our approach deals with a partial graph. Only Lemma 4.6 is a well-known fact about structural and observational dominance from existing work, and it cannot explain all these novel results.

"... more robustness than RCD is not supported: both drop 15 points."

RCA-DCM achieves 0.99 at λ=3\lambda=3 and 0.84 at λ=10\lambda=10, while RCD achieves 0.74 at λ=3\lambda=3 and 0.59 at λ=10\lambda=10. By "more robust", we meant that RCA-DCM is more accurate than RCD at both confounding strengths, not that it degrades less. We will fix the wording.

"... A small chain with a confounder would be more convincing here."

It is not quite clear to us what the reviewer meant by this. Would the reviewer mind clarifying how the X→Z→YX\to Z\to Y chain demonstrates the spurious shift propagation through unobserved confounders and the existence of equally plausible root-cause sets?


Q1. Scoring.

"Alg. 1 scores candidates with the training loss ... ranked by a Wasserstein distance ... Which is it? ... joint over VV, or per variable?"

We train the models with causal normalizing flows [4] by maximizing the likelihood, which fits the joint distribution over all variables. After training, we measure how far the distribution of our anomalous model is from the true anomalous data. We applied different distances, such as Wasserstein and MMD, and found the Wasserstein distance robust across setups. We will correct Alg. 1 accordingly.


Q2. Training.

"How is maximum likelihood done with shared latent noise ...? ... how exactly is fYf_Y shared?"

The three descriptions are equivalent; only the wording differs. Testing YY fits a pair (M^n,M^a)(\hat M_n,\hat M_a) to (Pn,Pa)(P_n,P_a) with fYf_Y tied and all other mechanisms free, i.e., 2∣V∣2|V| DCMs in total. Since M^n\hat M_n is the same for every candidate, we pre-train it once and, per candidate, fine-tune the pair jointly on Lnorm+LanomL_{norm}+L_{anom}. fYf_Y is the same module in both models, so it is updated by both losses. We use one normalizing flow per variable. For variables that share a latent, we sample random noise from N(0,I)N(0,I) and feed it to each of their flows, concatenated with the data input (Alg. 2). We will keep only this description in the revision.


Q3. Microservices.

"What is the 'service under test' ...? If it is the injected service, that leaks the label."

The "service under test" is the candidate being tested, not the injected service. Each injected fault is one problem instance. The injected service is the ground truth, is used only for evaluation, and is never an input.

"Please also report RCA-DCM with the true call graph ..."

Graph given to RCA-DCMSock Shop T@1Online Boutique T@1
Confounded sink (paper)0.8880.776
True call graph0.8880.776
True call graph plus shared latent0.8880.776

Q4. Inconsistent numbers.

"The abstract reports PRR 0.846, while Fig. 4(c) ... report 87% ..."

The chamber has 52 cases, so PRR moves in steps of 1/52. The abstract's 0.846 (44/52) comes from a run with early stopping, and the text's 0.865 (45/52) from a run with a fixed 50 epochs.

AuthorsOpenReview ↗

Rebuttal Part 3/3

Q5. NSigma and BARO in the chamber.

"Why do NSigma and BARO reach only 6% ...? Is the intervened variable ... dropped as constant?"

No: the target is present in all 52 cases. Both methods score each variable VV by its largest signed standardized deviation over the anomalous window, s(V)=max⁡t(vt−μn(V))/σn(V)s(V)=\max_t (v_t-\mu_n(V))/\sigma_n(V). Most perturbed variables are settings that are constant in the normal data, so σn(V)=0\sigma_n(V)=0; their implementation falls back to σn(V)=1\sigma_n(V)=1, giving s(V)=ca−cns(V)=c_a-c_n. The intervention usually lowers the setting, so s(V)<0s(V)<0 and the root cause is not at the top. When the root does vary normally, its children score just as high, and without a graph neither method can tell them apart. We also added a no-fault control (two disjoint splits of the reference data): every true root cause scores above the 95th percentile of the no-fault scores, while only a small fraction of non-roots do.

Finally, we are happy that the reviewer described our work as "a way to go beyond CI-based RCA under hidden confounding" and found our method consistently strong against the baselines. We highly appreciate that the reviewer found our work a useful practical tool if the concerns are resolved. We believe we have addressed all major concerns, and we will fix the minor concerns in the camera-ready version. We are open to further discussion. If the reviewer thinks we have resolved their major concerns, we hope they will increase our score to acceptance.

References

[1] CausalGAN: Learning causal implicit generative models with adversarial training.
[2] The causal-neural connection: Expressiveness, learnability, and inference.
[3] Modular learning of deep causal generative models for high-dimensional causal inference.
[4] Causal normalizing flows: from theory to practice.

Reviewer_PsX8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies root cause analysis under latent confounding. RCA-DCM formulates root-cause identification as testing whether the normal and anomalous distributions can be explained by an SCM while keeping a candidate mechanism invariant. In practice, the method fits deep causal models with each candidate mechanism shared across the two regimes and ranks candidates by the resulting fitting discrepancy. The paper provides theoretical analysis of feasible root-cause sets and graph misspecification, and evaluates the method on synthetic data, Causal Chamber, SockShop, and Online Boutique.

Critical strengths

The paper addresses the practically relevant problem of RCA with latent confounders and provides a flexible DCM-based implementation that can capture constraints beyond conditional independence.

The evaluation covers synthetic, physical, and microservice benchmarks, with generally strong empirical performance.

Critical weaknesses
  1. The theoretical novelty and its connection to the implemented algorithm are unclear. Testing whether two environments are compatible with a causal model while keeping selected mechanisms invariant is closely related to existing work on mechanism-change/intervention-target identification and latent-variable causal compatibility. Moreover, the theory concerns oracle SCM feasibility, whereas the implementation ranks variables using the residual loss of trained neural DCMs. A high empirical loss can also arise from model approximation or optimization error, which is not covered by the main guarantees.

  2. The experiments do not cleanly isolate the benefit of handling latent confounding. In Causal Chamber, hiding the root cause simultaneously introduces latent confounding and changes the problem from single-root to potentially multi-root recovery, while some baselines cannot recover multiple roots by design. In SockShop and Online Boutique, the shared latent confounder used by RCA-DCM is an assumed model structure rather than ground truth. Thus, the current results do not clearly establish that the improvements specifically come from handling latent confounding.

Questions for authors
  1. What is the key theoretical result that is genuinely new relative to existing work on mechanism-change/intervention-target identification and latent-variable causal compatibility?

  2. How does the theory justify using the residual loss of a finitely trained neural DCM as a proxy for SCM infeasibility? Can optimization/model-capacity failure be distinguished from genuine incompatibility?

  3. Can the Causal Chamber experiment separate the effect of latent confounding from the effect of multiple-root recovery?

  4. How sensitive are the SockShop/Online Boutique results to the assumed shared latent confounder? In particular, what happens with alternative graph/confounding specifications?

3 comments
AuthorsOpenReview ↗

Rebuttal Part 1/2

We thank the reviewer for the clear summary and constructive review. We appreciate that the reviewer found the problem of RCA with latent confounders practically relevant, the DCM-based implementation flexible enough to capture constraints beyond conditional independence, and the evaluation broad, with generally strong empirical performance.


W1 / Q1. Theoretical novelty.

"Testing whether two environments are compatible with a causal model while keeping selected mechanisms invariant is closely related to existing work ..."

"What is the key theoretical result that is genuinely new ...?"

We agree with the reviewer. Representing the two environments as causal models and the anomaly as a mechanism shift is established [1,2,3]. These methods, however, rely on conditional independence tests under causal sufficiency. Unobserved confounders induce spurious dependencies that invalidate these tests, so the methods do not apply in our setting. Methods that allow latents, such as [4,5], use invariance or CI-type constraints, and under latent confounding they can return only a superset of the true targets, or they require more environments than the two available in RCA for better performance.

Our main novelty is the RCA test. To decide whether YY is a root cause, we allow every mechanism except fYf_Y to shift and ask whether any pair of causal models compatible with the graph can still produce (Pn,Pa)(P_n,P_a). If no such pair exists, YY must be a root cause (Thm. 4.13). The test is motivated by instrumental-variable testing: like the instrumental inequality, it uses constraints beyond conditional independence, and it requires neither CI tests nor causal sufficiency. To our knowledge, we are the first to utilize distributional constraints beyond CI tests, such as the instrumental inequality (through DCM training), for RCA. If the reviewer has existing work in mind, we would be grateful if they could share it with us.


W1 / Q2. Oracle SCM feasibility.

"... the theory concerns oracle SCM feasibility, whereas the implementation ranks variables using the residual loss of trained neural DCMs."

Our theory does not require access to the true SCM. Eq. (2) asks whether any SCM pair compatible with the graph reproduces (Pn,Pa)(P_n,P_a), so it suffices to search over proxy models that induce the same observational distributions. A DCM that is representative enough for the ADMG (Def. 4.11) can realize every distribution induced by an SCM on that graph [6,7,8]. The infimum over DCM pairs therefore equals ℓ∗\ell^{\ast}, and only finite-sample, approximation and optimization error remain.


W1 / Q2. Approximation and optimization error.

"A high empirical loss can also arise from model approximation or optimization error ..."

In this paper, we assume perfect model training and address only the challenge of low sample size, since approximation and optimization error can be reduced by training an arbitrarily large model for many epochs, or by using a tabular foundation model. For the sampling error, App. E gives:

Proposition E.1 (Correctness of ranking). Suppose model training is perfect. If min⁡Y∈R∗ℓ∗(Y)>2ϵstat\min_{Y\in R^{\ast}}\ell^{\ast}(Y)>2\epsilon_{stat}, then with probability at least 1−δ1-\delta, min⁡Y∈R∗ℓ^(Y)>max⁡Z∉R∗ℓ^(Z)\min_{Y\in R^{\ast}}\hat\ell(Y)>\max_{Z\notin R^{\ast}}\hat\ell(Z).

Intuitively, every root cause is scored strictly above every non-root cause, so the k=∣R∗∣k=|R^{\ast}| root causes occupy the top kk positions of the ranking returned by RCA-DCM.

In practice, neural network models generally have some approximation error when matching the training distribution, so even a non-root candidate has a positive loss. Our ranking nevertheless remains sound, as Prop. E.1 shows: its correctness depends on the gap between the losses of root and non-root candidates, not on any loss being exactly zero.

The result extends trivially to approximation and optimization error. Our ranking is correct if min⁡Y∈R∗ℓ∗(Y)>2ϵstat+max⁡Z∉R∗[a(Z)+o(Z)]\min_{Y\in R^{\ast}}\ell^{\ast}(Y)>2\epsilon_{stat}+\max_{Z\notin R^{\ast}}[a(Z)+o(Z)], where a(Z)≥0a(Z)\ge0 is the approximation error of the DCM class and o(Z)≥0o(Z)\ge0 the optimization error for a non-root candidate ZZ. Only non-roots appear on the right: since ℓ∗(Z)=0\ell^{\ast}(Z)=0, a non-root's score consists of these errors alone, which can lift it above a root, whereas the same errors on a root only increase its score. This condition is not specific to our approach: for any ML method, a root cause cannot be detected unless its signal exceeds the noise.

AuthorsOpenReview ↗

Rebuttal Part 2/2

W2 / Q3. Isolating latent confounding from multi-root recovery.

"... hiding the root cause simultaneously introduces latent confounding and changes the problem from single-root to potentially multi-root recovery ..."

We would like to humbly point out that latent confounding and multiple root causes are not two independent challenges. Consider a system with a single root cause YY and a latent confounder between YY and its child, i.e., Y→Ch(Y)Y\to Ch(Y) and Y↔Ch(Y)Y\leftrightarrow Ch(Y). The non-root Ch(Y)Ch(Y) may shift in the anomalous data through the confounding path Y↔Ch(Y)Y\leftrightarrow Ch(Y), but since YY is the root cause, it typically shifts more. Even a method that only measures the median shift then ranks YY above Ch(Y)Ch(Y), which is the correct ranking. Latent confounding becomes a challenge when there are multiple root causes R1,R2R_1,R_2: we then test whether R1,R2R_1,R_2 occupy the top two positions, or whether a non-root appears between them, as shown in the introduction (Fig. 1(b)).

"... some baselines cannot recover multiple roots by design."

The baselines we compare with are compatible with multi-root recovery. Every synthetic experiment has ⌊p/2⌋\lfloor p/2\rfloor root causes: three in Exp. 1 (Fig. 4(a)) and up to five in Exp. 2 (Fig. 4(b)). The baselines recover these sets, not perfectly but well enough (at λ=3\lambda=3: BARO 0.91, NSigma 0.86, RCG 0.85, RCD 0.74, CIRCA 0.62).


W2 / Q4. The assumed graph for Sock Shop and Online Boutique.

"... the shared latent confounder used by RCA-DCM is an assumed model structure rather than ground truth ..."

The microservice graph is an overstated assumption, which we see as a strength. For Sock Shop and Online Boutique we use no ground-truth graph. Since RCA-DCM needs a graph, we supply an overstated one: every service points to the candidate, and a shared latent confounds all services. Assuming this graph rather than the ground truth means that we use less information. Even so, RCA-DCM attains the highest top-1 accuracy on both benchmarks. This shows that the method applies when the true graph is unavailable. Replacing the assumed graph with alternative specifications leaves the results essentially unchanged:

Graph given to RCA-DCMSock Shop T@1Online Boutique T@1
Confounded sink (paper)0.8880.776
Plain sink, no latent0.8800.776
True call graph0.8880.776
True call graph plus shared latent0.8880.776

Where the gains come from. The gains come from the test itself: only fYf_Y is held invariant while every other mechanism may shift. This design handles latent confounding, and as a byproduct it also performs well without confounders (Exp. 2, Fig. 4(b): best PRR at both sizes, 0.98 and 0.54).

We would be happy to have further discussion. If we have addressed your concerns, we kindly request that you consider increasing your score.

References

[1] Root cause analysis of failures in microservices through causal discovery.
[2] Causal inference-based root cause analysis for online service systems with intervention recognition.
[3] Root cause analysis of failures from partial causal structures.
[4] Causal discovery from soft interventions with unknown targets: Characterization and learning.
[5] Learning unknown intervention targets in structural causal models from heterogeneous data.
[6] CausalGAN: Learning causal implicit generative models with adversarial training.
[7] The causal-neural connection: Expressiveness, learnability, and inference.
[8] Modular learning of deep causal generative models for high-dimensional causal inference.

Dear Reviewer PsX8,
Thanks again for you effort on reviewing our RCA-DCM paper. We hope our rebuttal has addressed the concerns you raised. If any point remains unclear, we would be glad to address them in the remaining time.

If you feel your concerns have been addressed, we would greatly appreciate it if you could increase your score. Thank you.

Reviewer_aK2COpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper reframes root cause analysis as detecting which mechanisms must have changed to explain the distribution shift from normal to anomalous. It defines a solution set SolRC(G, Pn, Pa) over a given ADMG, proves it is upward closed (Lemma 4.8), and shows that any candidate whose exclusion leaves the distributional difference unexplained must be a root cause (Thm 4.13); the proposed RCA-DCM fits a pair of deep causal models per candidate and ranks by the loss of the fitted pair to estimate this. Experiments include two synthetic settings, the Causal Chamber with the root cause observed and hidden, and Sock Shop and Online Boutique against five baselines. The solution-set formalization is a genuine contribution; my problems are with what the paper reports and what it actually tests.

Critical strengths
  1. It defines "which sets of mechanism changes can explain the data" as an explicit mathematical object, instead of hiding it inside a CI test.
  2. The choice of datasets is fairly comprehensive, covering both simulated and real data, and the runtime cost is honestly reported.
  3. It pins down the direction of the bias in the output under graph misspecification, which is genuinely useful for anyone who only has a topological order.
Critical weaknesses
  1. App. G.1 logs MMD and median shift as "diagnostics". Re-ranking candidates by median shift is close to the failure mode identified in the introduction, so what would the same candidates score under such a re-ranking?
  2. How was the graph in the chamber hidden-root-cause experiment constructed? If what was fed in is a bidirected clique, then by the paper's own account the true root-cause set is that clique itself, and circling the clique is already a strong baseline.
  3. In the chamber's observed condition only one sensor is deliberately perturbed at a time, yet BARO and NSigma score 0.06; there is no no-fault control, so it is impossible to tell whether the methods fail or the scoring does.
  4. Alg. 1, the Reproducibility Statement, and App. G.1/G.6 describe inconsistent training procedures, and the gradient is attributed to two different losses. Which procedure and which loss produced the reported numbers?
Questions for authors

See weakness

7 comments
AuthorsOpenReview ↗

Rebuttal Part 1/2

We thank the reviewer for the careful review. We appreciate that the reviewer found the solution-set formalization a genuine contribution, the choice of datasets comprehensive with honestly reported runtime, and the direction of bias under graph misspecification genuinely useful.


W1. Re-ranking by median shift.

"Re-ranking candidates by median shift is close to the failure mode identified in the introduction, so what would the same candidates score ...?"

The same candidates score much lower: top-1 accuracy drops from 0.888 to 0.432 on Sock Shop and from 0.776 to 0.392 on Online Boutique.

The key difference is what is compared. Marginal-shift methods (BARO, NSigma) compare raw normal with raw anomalous data and ask "how much did YY move?". A downstream symptom moves a lot even though its own mechanism is intact, which is the failure mode of Sec. 1. We instead compare the true anomalous data with data generated by our model in which fYf_Y is held invariant across regimes, and ask "can the system explain the anomaly without changing YY?". A symptom passes this test, because its parents carry the shift; a root cause fails it. Wasserstein, MMD and median shift are three ways to measure this misfit. We rank by Wasserstein and log the other two only as diagnostics.

Score (generated vs. true anomalous)Sock Shop T@1 / T@3Online Boutique T@1 / T@3
Wasserstein (reported)0.888 / 0.9920.776 / 0.896
MMD0.320 / 0.7040.464 / 0.808
Median shift0.432 / 0.6800.392 / 0.632

Median shift sees only changes in location, so it misses misfits in spread and shape; the Wasserstein distance captures all of them. We will add this ablation to the main text.


W2. Graph in the hidden-root-cause chamber experiment.

"How was the graph in the chamber hidden-root-cause experiment constructed? If ... a bidirected clique, ... the true root-cause set is that clique itself ..."

When the root cause UU is hidden, the causal chamber graph becomes an ADMG over the observed variables: directed edges A→BA\to B, and a bidirected edge A↔BA\leftrightarrow B for every pair of children of UU (A←U→BA\leftarrow U\to B). This ADMG is the input to the algorithm. Since UU is unobserved, its children are the root causes, so the bidirected edges form a clique among them, as the reviewer notes. The clique is therefore not a baseline but the ground-truth solution R∗R^{\ast}.

This correspondence, however, is never given to any method. Each method must decide which variables are root causes from the observed data and its input graph; nothing in the input says whether they lie inside the clique or elsewhere. The experiment is therefore sound.

AuthorsOpenReview ↗

Rebuttal Part 2/2

W3. BARO and NSigma in the chamber, and a no-fault control.

"... only one sensor is deliberately perturbed at a time, yet BARO and NSigma score 0.06; there is no no-fault control ..."

The low scores come from how these two methods are designed, not from our pipeline. Both score each variable VV by its largest signed standardized deviation over the anomalous window, s(V)=max⁡t(vt−μn(V))/σn(V)s(V)=\max_t (v_t-\mu_n(V))/\sigma_n(V), and rank variables by s(V)s(V). Here μn,σn\mu_n,\sigma_n are the mean and standard deviation of the normal data for NSigma, and the median and interquartile range for BARO. This score suits microservice metrics, where a fault raises a metric such as CPU or latency. The chamber breaks it in two ways:

  1. Most perturbed variables are settings, not sensors. A setting is constant in the normal data, so σn(V)=0\sigma_n(V)=0 and the standardization is undefined. Their implementation falls back to σn(V)=1\sigma_n(V)=1, so the score becomes the raw change of the setting, s(V)=ca−cns(V)=c_a-c_n, where cnc_n and cac_a are its normal and anomalous values. The intervention usually lowers the setting (ca<cnc_a<c_n), so s(V)<0s(V)<0 and, when variables are ranked by s(V)s(V), the true root cause is not at the top.
  2. When the root cause does vary normally, its children score just as high. A child CC inherits the shift, so s(C)s(C) is as large as s(V)s(V), and the maximum over the window lets even unaffected sensors reach a large ss by chance. Without a graph, neither method can tell a root cause from its child. Since the two methods differ only in the choice of (μn,σn)(\mu_n,\sigma_n), they fail on the same cases.

No-fault control. We added one by running RCA-DCM on two disjoint splits of the reference data, where no mechanism changes. Since there is no root cause, these runs show the scores variables get from noise alone. We then compare the real fault cases against them: every true root cause scores above the 95th percentile of the no-fault scores, while only a small fraction of non-root variables do. Our score therefore separates faults from noise.


W4. Training procedure and loss.

"... describe inconsistent training procedures ... Which procedure and which loss produced the reported numbers?"

The procedures are equivalent; only the wording differs. Testing YY fits a pair (M^n,M^a)(\hat M_n,\hat M_a) to (Pn,Pa)(P_n,P_a) with fYf_Y tied and all other mechanisms free, i.e., 2∣V∣2|V| DCMs in total. Since M^n\hat M_n is the same for every candidate, we pre-train it once and, per candidate, fine-tune the pair jointly on Lnorm+LanomL_{norm}+L_{anom}. This warm start is an efficiency choice; the objective is still Eq. (2). Alg. 1 shows the plain version, and App. G.6 the warm-started one.

Def. 4.11 follows the original DCM framework (one network per node), which used a discriminator for image data. We follow the causal normalizing flow implementation [1], which trains by maximum likelihood. During evaluation, we measure how well M^a\hat M_a fits the anomalous data by comparing its samples with the data; we found the Wasserstein distance robust across setups.

We will make Alg. 1, Def. 4.11, App. G.1, G.3 and G.6 consistent with this description.

We would be happy to have further discussion. If we have addressed your concerns, we kindly request that you consider increasing your score.

References

[1] Causal normalizing flows: from theory to practice.

Reviewer_aK2COpenReview ↗

Re to Rebuttal 1/2

Thank you for the clarifications and additional experiments. They address some of my concerns, but a few questions remain.

W1. The new results show that Wasserstein performs better than median shift on both datasets. This resolves my concern.

W2. Since the bidirected clique matches the ground-truth root-cause set, simply identifying the clique members could already be a strong baseline. I would like to see a comparison with this simple baseline to better understand what RCA-DCM adds.

Comments on the paper

2 comments
Area_Chair_94YHOpenReview ↗

related literature

One additional question from my side after reading abstract and a first glance at the paper:

RCA with latent variables is considered also in Schkoda and Janzing “Root cause analysis in unknown cyclic graphs”, AISTATS 2026. While most of it is linear, it also discusses a non-parametric version in Appendix B. Appendix D.5 also contains an experiment with non-linear relations. Note that it is supposed to work even with a single sample from the anomalous regime. Please check relatedness and whether this requires rephrasing novelty and contributions or additional comparisons (which is a bit late for rebuttal I admit). Sorry for stepping in so late.

AuthorsOpenReview ↗

Relation and Comparison to Schkoda & Janzing (AISTATS 2026)

Dear AC,
Thank you for taking the time to look into our paper and for pointing us to this work. Yes, we agree this paper is closely related. We will cite it, and we have run a direct comparison.

Novelty. Our central claim is RCA in nonlinear models with latent confounders using constraints beyond CI, which is not fully addressed by their approach, since its guarantees rely on a linear structure.

Qualitative comparison. Both papers allow latent variables and model the anomaly as a sparse mechanism shift. The differences:

  • Mechanisms: they assume a linear SEM (cycles allowed); RCA-DCM is nonlinear and acyclic.
  • Graph: they do not need a graph; we need a (partial) ADMG, or assume a confounded sink graph (Exp 5) when none is available.
  • Anomalous data: a single sample is enough for them; we need an anomalous dataset.
  • Runtime: their method only estimates a precision matrix, with no neural network training, and is fast. RCA-DCM trains deep causal models and is comparatively slower, although candidates run in parallel (App. G.6).

App. B. App. B restates their score as a test of Y_j ⊥ F | Y_−j and notes that this holds without linearity. However, no nonparametric algorithm is given; the authors say estimation "is complex" and leave it open. As a CI test it has the failure mode demonstrated in the introduction of our paper (Fig. 1(a.ii)): Y depends on F given X, even though Y is not a root cause.

App. D.5. Based on our understanding, App. D.5 runs the linear method on a mostly-additive SEM, so it can a robustness check and might not be a nonlinear method.

Empirical comparison. We ran the baseline authors' unmodified code (https://github.com/DanielaSchkoda/CyclicRCA) on the same cases as RCA-DCM in three setups: our Exp 1 with increasing confounding (λ = 3, 10), and the Sock Shop and Online Boutique microservice datasets. We first reproduced their published RCAEval table. When several anomalous samples are available, we combine their scores by the maximum, as in their Sec. 5.2. We also implemented a nonlinear version of App. B with gradient boosting.

Sock Shop T@1Online Boutique T@1Exp 1 λ=3 PRRExp 1 λ=10 PRR
RCA-DCM0.8880.7760.990.84
S&J (linear)0.7520.7200.670.49
S&J App. B, nonlinear (our implementation)0.7600.6720.850.63

We are happy to provide further comparative discussion. We will provide a deeper qualitative and quantitative discussion in the camera-ready version.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.