- 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 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 . This has to be clearer in the revised version.
The evaluation does not support the claims
Scale: Exp 1 has observed variables (+4 latents), and Exp 2 has , with only two points per axis. In Exp 2, PRR already falls from 98% to 54% between and . 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 , and .
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 for a non-sink then only forces , because the latent is rich enough to fit any joint distribution with that marginal. The test then reduces to a per-service comparison of '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 that is never named, then uses 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 per variable with a product distribution , which does not match shared confounders. is written three different ways.
- l.147. "mDAG" is used without a definition...
- Notation clashes. is both an SCM (Def 3.1) and a set of distributions (Def 4.2), so "" reads as a distribution belonging to an SCM. denotes a neural network (Def 4.11, Alg. 2) while 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 instead of .
- l.167. Point "Graph 1" to Fig. 2(a).
- Fig. 2 leaves the bidirected arcs unlabelled, while Fig. 1(a) labels . 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. ). A small chain with a confounder would be more convincing here.
- l.287-291. "absence of the edge " and "R must be a root cause" should use . 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 ").
- 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 :
-
Scoring: Alg. 1 scores candidates with the training loss . 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 , or per variable? If per variable, joint constraints such as the IV one cannot affect the score at all.
-
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 shared?
-
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.
-
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?
-
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
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 and 0.84 at , while RCD achieves 0.74 at and 0.59 at . 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 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 , 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 shared?"
The three descriptions are equivalent; only the wording differs. Testing fits a pair to with tied and all other mechanisms free, i.e., DCMs in total. Since is the same for every candidate, we pre-train it once and, per candidate, fine-tune the pair jointly on . 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 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-DCM | Sock Shop T@1 | Online Boutique T@1 |
|---|---|---|
| Confounded sink (paper) | 0.888 | 0.776 |
| True call graph | 0.888 | 0.776 |
| True call graph plus shared latent | 0.888 | 0.776 |
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.
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 by its largest signed standardized deviation over the anomalous window, . Most perturbed variables are settings that are constant in the normal data, so ; their implementation falls back to , giving . The intervention usually lowers the setting, so 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.
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=D1∑d=1D10∣Top10(d)∩Rd∗∣, where D=30
is the number of datasets, Rd⋆ the true root-cause set of dataset d, and Top10(d) the method's top 10 nodes for that dataset:
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-FCI (Jaber et al., 2020) and LIT (Yang et al., 2024):
Ψ-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 minY∈R⋆ℓ⋆(Y)>2ϵstat+maxZ∈/R⋆[a(Z)+o(Z)], where a(Z),o(Z)≥0 are the approximation and optimization errors of a non-root Z. Small anomalous windows increase ϵstat, 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∣ tests for ∣V∣ variables independently. Whenever we test a variable Y, we assume it is a sink node: a confounded sink, in which every other service points to Y 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.