acceptodds
Under review as a conference paper at ICLR 2027

Causal-Structure-Preserving Transformation for Root-Cause Discovery

Abstract

Root-cause discovery (RCD) for a single interventional sample is important in many applications, including rare-disease diagnosis and cloud-service monitoring. Existing methods for one-sample RCD typically assume a linear structural equation model (SEM). In practice, however, observed common causes may affect variables through nonlinear or interaction effects, such as those involving age, gender, and experimental batch in RNA-sequencing studies, or incoming request load in cloud services. Ignoring these common causes can result in unaccounted-for confounding, whereas directly incorporating their nonlinear or interaction effects violates the linearity assumption underlying existing methods. These complications can obscure the sparse intervention signal and cause existing RCD methods to misidentify the root cause. To address this problem, we propose CSP-RCD, a two-stage procedure that first transforms the variables using flexible regression methods and then applies one-sample RCD algorithms to the transformed variables. We prove that this transformation preserves both the causal coefficient matrix and the sparse structural intervention of the original SEM, thereby providing a theoretical basis for the validity of downstream RCD methods. We provide a complete characterization of valid additive transformations and establish consistency of root-cause recovery for CSP-RCD. The proposed transformation can also be applied to other causal discovery tasks. For example, it can be combined with LiNGAM to estimate the causal structure when the variables are influenced by nonlinear effects of observed common causes. Extensive experiments on synthetic causal systems and real-world applications to the PetShop microservice benchmark and an RNA-sequencing dataset show that our proposed approach improves root-cause discovery.

Primary area:
causal reasoning
Keywords:
Root-cause discovery, causal discovery, single interventional sample, non-linear common causes

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_TdNnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Existing one-sample Root-cause discovery(RCD) method typically assume a linear structural equation model(SEM) and may therefore fail to adequately account for nonlinear or interaction effects of observed common causes on the variables. To address this issue, the authors propose CSP-RCD, a two-stage procedure that first transforms the variables by estimating the conditional mean m(w)=E[X∣W=w]m(w)=\mathbb{E}[X\mid W=w] from observational data to remove common-cause contributions from both observational and interventional samples, and then applies existing RCD methods to the transformed variables. Under the stated assumptions, the authors show that the population-level transformation preserves both the causal coefficient matrix BB and the sparse intervention vector δ\delta of the original SEM, providing a theoretical basis for applying downstream RCD methods to the transformed data.

Critical strengths

S1. The method handles nonlinear common-cause effects without estimating BB or gg. Because m(w)=E[X∣W=w]=(I−B)−1{b+g(w)}m(w)=\mathbb{E}[X\mid W=w]=(I-B)^{-1}\{b+g(w)\}, removing the propagated common-cause contribution only requires regressing XX on WW, rather than separately estimating the structural coefficient matrix BB and the nonlinear function gg. The resulting two-step procedure applies existing one-sample RCD methods to the transformed data, so it extends their scope without designing a new root-cause scoring algorithm, and it preserves their property of not requiring the causal graph. The flexibility to choose the regression model for the conditional mean is also practical.

S2. The proposed transformation is theoretically justified and linked to the estimation procedure. The authors not only show that removing the conditional mean preserves BB and δ\delta, but also characterize additive transformations satisfying this property up to a constant difference. This helps explain why this specific transformation is appropriate and why other simple corrections fall short.

Critical weaknesses

W1. The stability regarding conditional mean estimation errors and practical application guidelines are not sufficiently clear. While the structure-preservation results hold for population-level transformations using the true conditional mean, the practical CSP-RCD relies on the estimated conditional mean. Consequently, if the regression is inaccurate especially around the common-cause values of single-intervention samples, the adjustment process may actually worsen root-cause estimation. This sensitivity is evident in Table 2. For instance, in Test13 under high-traffic conditions, the cause ranking is 12th on the raw data, 7th using the Random Forest-based transformation, and 19th using the GAM-based transformation.

Questions for authors

Q1. How were the nuisance learners and their hyperparameters (especially the random forest on PetShop) selected, and was this done using only observational data, without access to the true root causes? Were the transformed observational samples computed with in-sample or cross-fitted predictions? Finally, was cross-validated error useful for identifying cases where the transformation worsens the rank (e.g., GAM in Table 2(a), Test11–13)

Q2. In choice of WW in PetShop, how is the non-descendant assumption justified at the level of the measured latency variables, given that an incident could affect these services through shared load or retries? Are there incidents whose true root cause lies in the selected WW, and if so, how were they handled in the evaluation?

Reviewer_Jn81OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper talks about a two-stage procedure that first transforms the variables using flexible regression methods and then applies traditional RCD algorithms to the transformed variables.

Critical strengths
  1. The paper is written clearly.
Critical weaknesses
  1. The major concern is method novelty: The proposed approach is a standard residualization well known and widely used by practitioners to adjust for known confounding.
  2. Theory is also weak: the theorem assumes the dimension is fixed, while RNA applciations usually have more variables than the sample size. Thus, the asymptotic theory is of limited practical relevance.
Questions for authors
  1. Can you strengthen the theory to allow dimensionality to grow?
  2. What if there are discrete variables which are modeled by a generalized additive model?
3 comments

Dear Reviewer Jn81,

Thank you again for taking the time to review our paper.

We understand your major concern regarding novelty, as we also considered the connection to standard residualization when studying this problem. In our response, we explained why its familiar form still requires careful study for the task we consider and how our contributions go beyond simply applying residualization.

As the author–reviewer discussion period is nearing its end, we would greatly appreciate it if you could let us know whether our response addresses your concern or whether you still have any questions. We would be happy to provide further clarification or answer any follow-up questions before the discussion closes on November 18.

Thank you again for your feedback, time, and consideration.

Best regards,
The Authors

Thank you for the detailed rebuttal.

  1. I am not saying that residualization always works. The paper proposes a procedure largely coinciding with the standard practice of residualization. In the proposed model, using residualization for covariate adjustment is the default practice, and its validity is immediate from the reduced form. For example, in your RNA-seq data application, handling covariates using residualization is a standard protocol for data analysis:
  • Stegle et al. (2012), Nature Protocols 7(3):500–507.
  • Brechtmann et al. (2018), AJHG 103(6):907–917.

The example in the rebuttal lies outside your proposed model and, in fact, demonstrates the limitation of the proposed method. Thus, the rebuttal does not resolve my concern about its novelty.

  1. My Q2 asks whether you can go beyond standard residualization to handle the cases where X variables are of discrete or mixed type. This is not addressed or properly discussed.

I appreciate the author's rebuttal. But I remain unconvinced that this submission contains enough contributions and meets the quality for acceptance in ICLR.

I am not saying that residualization always works. The paper proposes a procedure largely coinciding with the standard practice of residualization. In the proposed model, using residualization for covariate adjustment is the default practice, and its validity is immediate from the reduced form. For example, in your RNA-seq data application, handling covariates using residualization is a standard protocol for data analysis: Stegle et al. (2012), Nature Protocols 7(3):500–507. Brechtmann et al. (2018), AJHG 103(6):907–917.

Thank you for your further response and clarification.

We were aware that residualization is a standard procedure for data analysis and is widely used in practice to adjust for known covariates. We have discussed this in the related work section and have added the two works you mentioned.

However, the problem and model considered in our paper differ from those in the related work, including the two papers you mentioned: we aim to identify the root cause of a single interventional sample under linear SEMs with nonlinear common-cause effects. It is unclear from these related works whether residualization is a valid transformation for the problem we consider, and we have not found any prior work studying this problem.

In particular, one of our findings is that consistent estimation of the conditional mean used in residualization is not necessary for preserving the causal structure and the intervention vector, as an additive constant bias is allowed. We have discussed our other contributions and findings in our first response to your comment, so we do not repeat them here.

The example in the rebuttal lies outside your proposed model and, in fact, demonstrates the limitation of the proposed method. Thus, the rebuttal does not resolve my concern about its novelty.

The example provided in the rebuttal is to make the point that residualization does not work in general for all tasks, and thus its validity requires careful study depending on the task, model, and downstream method.

This example must lie outside our considered model, because we have proved that the class of additive transformations, including standard residualization as a special case, is valid by showing that these transformations preserve both the causal structure and the intervention vector.

My Q2 asks whether you can go beyond standard residualization to handle the cases where X variables are of discrete or mixed type. This is not addressed or properly discussed.

Thank you for clarifying the question.

Yes, our method can handle cases where the variables in XX are discrete or of mixed type, as long as the linear SEMs described in (2) and (3) and the other stated assumptions hold. We do not require continuous errors or a particular parametric error distribution.

We believe this question also reflects the value of this type of work: by carefully studying a problem, we can clearly understand the conditions under which the method is valid, as well as its applicability and limitations.

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

This paper extends prior work on root cause discovery for individual interventional samples beyond the linear setting to allowing common cause to come in nonlinearly. Via a transformation, they make sure that the causal structure (matrix B) stays the same and reduces to the linear problem - so that they can use known algorithms for the linear case. Authors provide

Critical strengths

The paper is extremely well written and was very nice to read in this day and age - sentences following each other made sense and there was a nice logical sequence and at no point was there anything really unclear.

Critical weaknesses

None.

Questions for authors

General weaknesses: The authors could explain a little bit why on certain tests the transformation doesn't give that much gain compared to others (i.e. in low traffic test0, test14; in high traffic test 11 for example. Their method that allows them to extend to non-linear is not magic and rather straightforward, but it's clear the paper had to be written and the experiments had to be done.

Suggestions to clarify: In Figure 1 it could be clearer what the x vs. y axis plots mean (i.e. clearer labels - the idea of "rank distribution" isn't super clear when you know a ground truth root case) Please use different linestyles in Figure 1, perhaps also with different markers - I would highlight the transformed line with a different linestyle rather than color.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Author–Reviewer Discussion

Dear Area Chair,

As the discussion period is nearing its end, we would like to briefly summarize the discussion and highlight our response to Reviewer Jn81's main concern regarding novelty, as we have not yet received any further feedback from the reviewer.

We have incorporated Reviewer LEWV’s suggestions on figure labels and line styles and added a discussion explaining why the transformation provides smaller gains in certain real-data cases, as they suggested. We have also incorporated Reviewer TdNn’s suggestions, and the reviewer has confirmed that our responses address their concerns and maintains their positive assessment.

Reviewer Jn81 raised method novelty as their major concern, viewing our approach as an application of standard residualization. We have provided a detailed response and would like to highlight the following points:

  • The validity of residualization for adjusting for confounding depends on the task. Residualization does not automatically resolve all problems related to confounding. Its validity for a particular downstream task requires careful study. We have provided a concrete example in our response to illustrate this point.

  • Our task requires preservation of causal structure and sparse interventions. In our problem, a valid transformation must remove the common-cause contribution while preserving both the linear SEM structure and the sparse intervention. Its validity also depends on specific assumptions, including the non-descendant and overlap assumptions in Theorem 1.

  • We characterize all valid additive transformations, with standard residualization as a special case. Specifically, we characterize the complete class of additive transformations satisfying these preservation requirements.

  • We study estimation error and downstream consistency. We study how estimation error in the transformation affects the empirical moments of the transformed data and provide a consistency result for downstream root-cause discovery.

  • Our transformation also applies to causal discovery and inference tasks beyond root-cause discovery. It extends the applicability of linear-SEM methods to settings with nonlinear common-cause effects. Our LiNGAM experiments illustrate the advantage of this extension for causal structure learning.

These contributions go beyond simply applying residualization, and we have revised the manuscript to make them clearer. Our full response to Reviewer Jn81 provides the supporting arguments and also responds to their questions about high-dimensional settings and discrete variables.

We would appreciate your consideration of these clarifications when assessing the paper. Thank you very much for your time.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.