acceptodds
Under review as a conference paper at ICLR 2027

Online Root Cause Analysis via Self-Gating Dual-Branch Causal Mechanism Tracking

Abstract

Reliable microservice operation hinges on three telemetry-monitoring tasks: anomaly detection, root cause analysis (RCA), and drift adaptation. All three concern the causal mechanisms generating the stream and must run online, on data already containing the faults to flag. Existing methods sidestep this setting by 1) consuming an anomaly timestamp, 2) assuming a clean training period, or 3) trusting an upstream detector, so faults silently pollute the learned reference. We show this pollution is invisible to detection yet fatal to localization: the reference absorbs exactly the contaminated fraction of each shift, shrinking the localization margin while the detection contrast stays positive. We therefore propose GDRC, a self-gating dual-branch causal mechanism tracker. Each window flows through two branches over a shared causal graph. The invariant branch maintains the reference of normal mechanisms; the variant branch refits the window as a sparse mechanism shift. A self-gating sampler compares the two by a BIC-penalized likelihood ratio, yielding per-node gains and a window responsibility. This responsibility gates the reference update, requiring no labels, no clean period, and no upstream detector. One framework serves all three tasks: detection, localization, and drift are read off the gains, so no timestamp is needed and no module absorbs another's errors. Experiments on controlled streams and two real benchmarks show that achieves competitive root-cause accuracy. Run end to end on a stream, it raises its own alarm and ranks there nearly as well as at the true fault onset. The same gain absorbs real traffic-regime changes while catching the faults that follow. Code and results are in the supplementary material.

Primary area:
learning on time series and dynamical systems
Keywords:
Root Cause Analysis, Causal Discovery, Distribution Shift, Anomaly Detection

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

The paper proposes GDRC, an online framework with an invariant branch maintaining normal mechanisms and a variant branch modeling sparse mechanism shifts. A penalized likelihood comparison gates reference updates and supplies signals for detection, localization, and drift adaptation. Two propositions characterize reference contamination, while synthetic experiments and evaluations on PetShop and RCAEval RE1 examine the framework’s effectiveness.

Critical strengths
  • Formulating three monitoring tasks through shared mechanism-shift evidence offers a coherent alternative to treating them independently.
  • The paper provides proofs characterizing reference contamination and its asymmetric impact on detection and localization under stated assumptions.
  • Explicitly protecting the reference from anomalous updates addresses an important challenge in continuous monitoring.
Critical weaknesses
  • Both benchmarks involve injected faults rather than naturally occurring production incidents. Their telemetry therefore provides limited evidence of reliability under operational failures and evolving production conditions.
  • Although diagnostic experiments are provided, some explanations remain conjectural. For example, attributing both RE1 detection and localization deficits to graph quality requires direct controlled evidence on RE1; graph substitution results on PetShop do not establish that conclusion.
  • The paper does not clearly establish which joint objective improves over a pipeline combining strong detectors, RCA models, and drift methods. Existing streaming experiments lack this integrated comparator. Numerous secondary analyses consequently leave the central advantage unresolved. A matched comparison of localization success, delay, false alarms, and adaptation would be more decisive.
Questions for authors

Please refer to the Weakness section above.

Reviewer_voEBOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GDRC, a root cause analysis method on time series data where they assume that unlike other methods, the faults are not identified and informed to the algo. The paper identifies that in such settings care must be taken to differentiate windows of data as whether they contain anomalous samples or not, otherwise, the estimates thought to be inferred from assumed normal regime may lead to "reference pollution". So, GDRC passes each window of data through a invariant normal and abnormal variant branch. A BIC based criterion judges whether the window is normal. If it is, then such a window is used to update the normal parameters through exponential forgetting. On synthetic datasets they show that GDRC performs best across very high pollution levels. On real benchmarks, GDRC is the best method on PetShop but falls behind simple methods on RCAEval RE1, where only a causal graph is learned.

Critical strengths

The paper is extremely well written with a very lucid flow. I could follow most of the paper although there are some missing details.

The introduction is very well written.

The problem statement appears synthetic to me. Why would real data consist of 40% contamination? Also since even with polluted data anomaly detection accuracy is high, can we not simply use that as a filtering criterion to subset the real data on which to fit the normal causal mechanisms?

The experiments are very well explained, although there are missing recent baselines: IDI (https://arxiv.org/abs/2505.00930), Traversal and Smooth Traversal (https://papers.nips.cc/paper_files/paper/2025/file/ec87ccda449e853c2876093d682cf3a5-Paper-Conference.pdf), etc.

Critical weaknesses

I have several questions on the missing details in the paper:

  1. Why do you assume a fixed basis in line 213. If a linear model is preferrable, why not try self explaining Neural nets that are more general.
  2. Line 215 mentions "sufficient statistics". What are they?
  3. Line 241 to 258: lots of details appear to be missing:

a. How are the likelihoods computed?

b. Explain whay gjg_j means. Please provide more details for Eq. 1

c. Also Eq. 2. Suppose we consider a chain of N nodes n1 -> n2 -> ... -> nN. How would estimate for w vary with the chain length NN and the location of the root cause? I see a sum over N nodes. Does this mean that a lengthy chain naturally has a propensity to increase the w?

d. Why is Gaussian likelihood the right choice given particularly you assumed a fixed basis?

e. Why is ww the right routing score? Do you have intuition behind it? Or is it the case that you tried several alternatives and this choice worked well in practice?

  1. At line 262, ata^t is computed as a sum over the gains. Again explain how this would vary with the chain example above where root cause appears early vs later in the chain.

Experients:

  1. Please include comparisons with missing baselines.

  2. In Fig 2, what would happen if we exclude samples where anomaly is detected while estimating the coefficients and then ind root causes? I am asking this particularly because the detection AUC seems robust to cc and so can be used as a strong filter to select just normal samples.

  3. In line 393, why do you filter 52 testcases from petshop? If a ground truth fault exists, a missed detection should be considered as a failure to detect root cause and should not be removed from the testcases.

  4. Why was ranked correlation not included in the main paper? With that the proposed method barely wins in real data. In RE1 the proposed method significantly trails behind BARO. Does this mean that several hyperparameters/assumptions introduced in the algo make the method perform poorly on real datasets?

  5. Overall please provide more details in the real data experiments and explain the cases where the proposed method wins/trails behind in a crisp and lucid manner.

Questions for authors

Please address my comments and questions in the weakness section.

Reviewer_vmJ9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a root anomaly detection and cause analysis scheme for microservices. It maintains an invariant branch representing normal mechanisms as a reference and a variant branch fitting the window as a sparse mechanism shift. The core insight of this paper is that updating the reference with faulty observations can undermine subsequent localization even when detection performance remains strong. The major contribution claimed in this paper is that it unified anomaly detection RCA, and drift adaption. The validation show some benefits of gating on synthetic streams and competitive performance in some real-data settings.

Critical strengths

It shows a challenging problem with compelling observation that detection quality can conceal localization failure.

Critical weaknesses
  • The real-system evaluation does not clearly establish the proposed framework’s advantage over existing methods. In addition, The benefit of adaptive self-gating is not sufficiently ablated.
  • Whether correlation or causation can be identified by the framework needs further justification.
Questions for authors
  • Does the implementation requires acyclicity when constructing the correlation graph? If there exist cycles, will there be any changes in the methods?

  • The parent sets seem to be selected by correlation what assumptions justify interpreting GDRC’s likelihood gains as causal mechanism changes, and what evidence shows that the method distinguishes the true root cause from correlated downstream symptoms?

  • GDRC does not clearly show a clear advantage in real benchmarks compared to the baselines, especially the RE1 benchmark. Can you explain where the regression is from?

  • Can you provide controlled comparisons that isolate the benefit of the proposed model-comparison gate over simpler alternatives, such as a frozen baseline reference, or residual-based gating?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.