acceptodds
Under review as a conference paper at ICLR 2027

LoCo: Local Competitive Learning for Interpretable Multivariate Time Series Anomaly Detection

Abstract

In streaming data scenarios, multivariate time series anomaly detection is critical for system reliability. Beyond detection, operators require interpretable diagnostics to make informed decisions and repair faults efficiently. However, state-of-the-art detection models are often computationally heavy and opaque, and existing root cause analysis methods are confined to locating anomalous variables at the current timestamp, overlooking the temporal evolution of anomalies along causal chains. Inspired by physical phenomena and real cases, we introduce two design principles—Finite Temporal Horizon and Causal Horizon. These principles and the sparsity of fault-relevant causal connections motivate the design of LoCo, a lightweight interpretable framework based on Local Competitive Learning. LoCo performs online anomaly detection via forecasting residuals and, upon detection, traces root causes along the learned time-lagged causal matrix, from current anomalous variables back to earlier upstream sources. This time-lagged causal matrix not only captures causal influences among variables across different past time steps, but can also be compressed into a global causal graph to accommodate conventional causal discovery tasks. Theoretically, the recovered lagged predictive support coincides with the true structural causal support under identifiability assumptions. Extensive experiments across anomaly detection, root cause analysis, and causal discovery demonstrate that LoCo achieves competitive accuracy and computational efficiency, while providing end-to-end interpretable visualizations suitable for resource-constrained online deployment. The code is available at https://anonymous.4open.science/r/LoCo-491F/.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Anomaly Detection, Local Competitive Learning, Interpretable Anomaly Detection, Root Cause Analysis, Causal Discovery

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZPx6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LoCo, a lightweight framework for multivariate time-series anomaly detection, root cause analysis, and causal discovery. It introduces a top-k gradient competition mechanism to learn sparse time-lagged dependencies and uses the resulting model for forecasting-based anomaly detection and backward root cause tracing. Experiments across multiple benchmarks demonstrate competitive detection accuracy, computational efficiency, and interpretability.

Critical strengths
  1. The paper addresses a meaningful problem by integrating anomaly detection, root cause analysis, and causal discovery into a lightweight framework. The time-lagged root cause tracing provides an interesting perspective on anomaly propagation.

  2. The experimental evaluation covers multiple datasets and three related tasks. The results demonstrate competitive accuracy and computational efficiency, supporting the practical potential of the proposed framework.

Critical weaknesses
  1. The theoretical justification contains a fundamental inconsistency. Theorem 4.3 relies on nonzero population gradients for true causal parents at the true parameter. However, under the stated SVAR model and exogeneity assumption, the expected gradients should vanish for all historical inputs at the true parameter. This makes Assumption C.5 incompatible with the preceding assumptions and undermines the claimed support recovery guarantee.

  2. The central claim of time-lagged root cause tracing lacks direct quantitative validation. The RCA experiments evaluate conventional variable-level root cause identification, while the proposed temporal tracing capability is mainly illustrated through case studies. Consequently, the accuracy of recovered propagation paths and time lags remains unclear.

Questions for authors
  1. Could the authors clarify how Assumption C.5 can hold at the true parameter when the expected residual gradients vanish under the stated model assumptions? How does this affect Theorem 4.3 and the finite-sample support selection result?

  2. Could the authors provide quantitative evaluations of time-lagged root cause identification, including the accuracy of recovered propagation paths and temporal offsets, rather than relying primarily on variable-level RCA metrics?

1 comment

Q1 About Theory
We thank the reviewer for pointing out the problems in the original theoretical analysis. After rechecking the assumptions and theorems, we agree with the reviewer’s concerns. The original manuscript tried to explain how Top-k gradient competition recovers the true causal support set by comparing gradient magnitudes between true parents and non-parents. However, this argument is not rigorous.

Specifically, original Assumption C.5 required true parents to have a strictly positive population-gradient lower bound at w∗w^*. But by C.1–C.2, when the model converges to w∗w^*, the population gradient should be zero at all coordinates, including true-parent coordinates. So C.5 indeed contradicts the earlier assumptions. We also realize that the original claim—that early-training Top-k gradient selection can recover the true causal support set. In particular, if the support set is already basically fixed early in training, the original explanation of continuous gradient competition during training is not accurate enough.

Following the reviewer’s suggestions and our further analysis, we reorganized the theory in Section 4 and Appendix C. We no longer use the original support-recovery argument. Instead, we study the predictive dependency information contained in gradients. In the new Phase 1, as requested by Reviewer fX4S, we further link gradients to Granger causality. Through the new Gradient–Residual Identity, we show that, under the relevant conditions, model-parameter gradients relate to the covariance between prediction residuals and historical inputs. Thus, gradients can reflect the information historical variables provide for current prediction, giving a new theoretical basis for understanding the time-lagged dependencies learned by LoCo.

On this basis, we adjusted the role of the later theoretical analysis. The new theory no longer tries to prove that early-training Top-k gradient selection necessarily recovers true direct causal parents, nor does it describe gradient competition as a process that guarantees exact recovery of the true causal structure. Instead, we more cautiously ground LoCo’s theoretical explanation in the incremental predictive information reflected by gradients and its link to Granger causality. Accordingly, we revised the original statements about causal support recovery.

Q2 Time-lagged Evaluation
We agree that traditional variable-level RCA metrics alone cannot directly evaluate time-lagged root-cause tracing. We therefore added time-lagged RCA and CD experiments on synthetic datasets to quantitatively evaluate root-cause localization and recovery of time-lagged causal relations. Settings and results are in Appendix E.4 G.3 G.5 H.11. For the experiments and comparisons with other baselines, see our reply to Reviewer AJfy’s Q4.

Reviewer_AJfyOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes LoCo, a single-layer forecaster for multivariate time series anomaly detection. It is trained on normal data with a row-wise top-k mask on the gradients, so that each target only updates the lagged inputs with the largest gradient magnitude. A linear version and a kernel version (per-variable, per-lag random Fourier features with group-wise top-k) are given. The learned weight matrix is then reused for three tasks: anomaly detection from standardized forecasting residuals, root cause analysis by tracing back along the lagged weights to earlier upstream variables, and causal discovery by averaging over lags and antisymmetrizing into an N x N graph. The authors also give a theoretical analysis arguing that the top-k gradient update recovers the true lagged parents under SVAR assumptions, and they evaluate on nine AD datasets, SWaT and MSDS for RCA, CausalRivers for causal discovery, and a SWaT case study where an FIT401 anomaly is traced back to P501.

Critical strengths
  • Very lightweight model. A single linear layer with around 26K parameters is cheap to train and run.
  • Competitive AD results. LoCo is best or second best on most of the nine datasets against much larger transformer and foundation-model baselines.
  • Interpretable by design. The learned weights can be read directly as lagged dependencies between variables, without any post-hoc explanation method.
  • Relevant problem setting. Tracing an alarm back to an earlier upstream variable and time lag, instead of just the most deviated variable at alarm time, is a real need in industrial monitoring.
Critical weaknesses
  • Theory is inconsistent. Assumptions C.1–C.2 make the population gradient at w* zero for every coordinate, which contradicts the gradient gap in C.5. Theorem 4.3 and Proposition C.7 therefore give no support recovery guarantee, and the claim in the abstract that the recovered support coincides with the true causal support is not supported.

  • The authors show (H.5) that the support is fixed in the first 8 epochs, far from w*, where gradients reflect marginal correlations and not conditional dependence. The "Causal Horizon" argument therefore does not explain how edges are actually selected.

  • Eq. 12 only masks the update direction and never projects the weights, so all coordinates updated during training stay non-zero, which Appendix C.6 itself admits. This contradicts the claim that a sparse causal topology emerges, and the sign mapping for Kernel-LoCo (Eq. 30) depends on the random RFF draw and has no causal meaning.

  • Eq. 7 measures how much the forecast changes when a parent is replaced, not how much of the residual it explains. Since e² − (e + c)² = −2ec − c², replacing an abnormal parent can even increase the error.

  • Algorithm 2 does not match Section 3.4. The text says the current node becomes a root cause when its parents are normal, but Algorithm 2 adds the normal parent instead, and it is not explained how ancestors beyond the L-step window are attributed.

  • In the SWaT case study, the top root cause (P501, t−11) removes only 57.38 out of an SPE of 6856.9, i.e. less than 1% of the anomaly, while the text concludes that the FIT401 anomaly "is caused by" P501. The negative cumulative scores in Table 12 are not possible under Algorithm 2, and the result is backed by a physical explanation rather than the SWaT attack labels.

  • Gain does not come from the proposed mechanism. In Fig. 6(c), top-k = 100% (a plain linear forecaster) is close to the best setting and would still beat most baselines. Comparisons to full-gradient training, L1-VAR and IHT under the same pipeline are missing.

  • Main novelty is not evaluated; temporal RCA and lagged causal discovery, the main novelty, are only shown in a case study. Appendix E.3 states that CausalRivers provides propagation lags, and no synthetic SVAR experiment with known lags is given.

  • RCA is compared only against ε-Diagnosis and AERCA. Strong recent baselines are missing, such as ShaTS[1] and CondAttr[2], and causal RCA methods like RCD[3], CIRCA[4], or BARO[5].

  • With max(C − Cᵀ, 0), a negative effect j → i becomes i → j, and feedback loops are always cut to one direction. It is also unclear whether VAR received the same post-processing.

  • No validation split or hyperparameter selection is described, and the best values in the sensitivity plots on test AUROC match the main tables. The Creditcard split (train and test both equal the full dataset size) and the ε-Diagnosis AC@3 on MSDS (0.0149, below the minimum AC@1/3 = 0.05 allowed by Eq. 68) also need explanation.

  • The Fig. 2 caption describes four panels while the figure has five, and the math labels inside the figure are not properly formatted. Appendix H.5 and H.10 are almost word for word the same, there are broken references, and Fig. 4(c) is labeled P502 instead of P501.

References:

[1] De La Peña, Manuel Franco, Ángel Luis Perales Gómez, and Lorenzo Fernández Maimó. "ShaTS: A shapley-based explainability method for time series artificial intelligence models." Future Generation Computer Systems (2025): 108178.

[2] Mishra, Shashank, et al. "Conditional Attribution for Root Cause Analysis in Time-Series Anomaly Detection." Joint European Conference on Machine Learning and Knowledge Discovery in Databases. Cham: Springer Nature Switzerland, 2026.

[3] Ikram, Azam, et al. "Root cause analysis of failures in microservices through causal discovery." Advances in Neural Information Processing Systems 35 (2022): 31158-31170.

[4] Li, Mingjie, et al. "Causal inference-based root cause analysis for online service systems with intervention recognition." Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining. 2022.

[5] Pham, Luan, Huong Ha, and Hongyu Zhang. "Baro: Robust root cause analysis for microservices via multivariate bayesian online change point detection." Proceedings of the ACM on Software Engineering 1.FSE (2024): 2214-2237.

Questions for authors
  • How can Assumption C.5 hold when C.1–C.2 imply that the population gradient at w* is zero for every coordinate, including the true parents?
  • The support is fixed in the first few epochs, when the model is far from w*. What guarantees that the top-k gradients at that point select the direct parents and not just strongly correlated variables?
  • Eq. 12 never projects the weights onto a sparse set. In what sense is the final W sparse, and how many non-zero entries per row does it have after training?
  • What is the AD, RCA and CD performance with top-k = 100% (no gradient competition), and how does it compare to L1-VAR and IHT under the same pipeline?
  • Which procedure produced the RCA results: the stopping rule in Section 3.4 or Algorithm 2? How do you get negative cumulative scores in Table 12 when Algorithm 2 multiplies absolute values?
  • Does P501 match the attack label for anomaly #2303 in the SWaT ground truth? If normalizing P501 removes less than 1% of the SPE, why should it be considered the root cause?
  • Why are the lags in CausalRivers (Appendix E.3) not used to evaluate the time-lagged graph?
  • How do you justify antisymmetrization when it turns a negative edge j → i into i → j? Was the same step applied to VAR?
  • How were L, top-k, D and σ selected? Was a validation split used, or were they picked from the test results shown in the sensitivity plots?
  • Comparing against stronger RCA baselines such as ShaTS, CondAttr, RCD, CIRCA or BARO would make the RCA results more convincing.
  • The Creditcard train and test sets are both listed with 284,807 samples, which is the size of the full dataset. Is this a typo, or do the two sets overlap? How can ε-Diagnosis have AC@3 = 0.0149 on MSDS when Eq. 68 requires at least AC@1/3 = 0.05?

Suggestions:

  • A synthetic SVAR experiment with known lagged graphs and known fault onsets would directly test the temporal RCA claim.
4 comments

We thank the reviewer for the constructive comments. We revised the paper accordingly; all changes are marked in yellow.

Q1 & Q2 Theory Inconsistency and Wrong Claims
We rechecked the assumptions/theorems and agree with the theoretical issues. Original Assumption C.5 required true parents to have a strictly positive population-gradient lower bound at w*. But by C.1–C.2, at true parameters all coordinates should have zero population gradient. So C.5 contradicts earlier assumptions, and the support-recovery guarantee lacks sufficient basis.

Following the comment, we reorganized Section 4 and Appendix C. We no longer use the support-recovery argument; instead we study predictive dependence in gradients. More details are provided in Reviewer ZPx6's Q1

Q3 Weight Sparsity
We agree: original GHT uses a Top-k mask to limit coordinates updated each time, but does not directly set other weights to zero. Thus sparse gradient updates ≠ strict sparsity of final weight matrix. With Q1/Q2 revisions, we clarify the gradient-competition mechanism and adjust algorithm description/claims. Appendix H.11 adds early-training analysis showing LoCo's causal structure before prediction error fully converges.

Q4 More Baselines and No Mask
Appendix H.11 adds time-lagged RCA and CD experiments on synthetic datasets and compares methods under a unified pipeline. LoCo performs overall well on time-lagged root-cause localization and causal graph recovery. Ablation with Top-k=100% shows removing the gradient mask brings no increase and lower topk already can recover the causal structure. Settings/metrics are in Appendix E.4, G.3, G.5.

Q5 Algorithm 2 & SPE increment
Section 3.4 and Algorithm 2 describe the same process: text vs formal algorithm, has clarified. Table 12 “Cum.” is signed cumulative contribution along the propagation path, Cum.=∑ci←pCum.=\sum {c_{i←p}}, so it can be negative. Original Algorithm 2 wrote path score as a product of absolute values, inconsistent with actual results; corrected and defined. Final ranking uses counterfactual $Attr(p^*)=|SPE(t)−SPE^{(p^{*})}(t)| $, not Cum. Algorithm/text unified.

And about the scenario of SPE increment in Eq. 7, we have discussed in reply to ReviewerfX4S'Q3.

Q6 SWaT Case

In Fig. 4(c), P502 is a typo and should be P501. In case #2303, the abnormal variable is FIT401, not P501. Since SWaT lacks complete ground-truth propagation paths, our P501–FIT401 analysis is based on the original paper and actual system structure.

SPE measures the whole system prediction error, while Attr(P501, t−11) measures its change after replacing P501's historical input. They represent different quantities and should not be directly compared.

SWaT has 51 variables (the average of each var.≈140\approx 140). Algorithm 2 selects the most deviant variable as the starting point for backtracking. Here, it is FIT401, with a standardized squared error of about 400. Other variables, maybe FIT402, have second largest errors but are not used as separate starting points. Thus, P501's attribution reflects its influence on FIT401, not the entire system anomaly. It already accounts for a substantial part of FIT401's squared error, and its influence may span multiple past time steps.

Q7 CausalRivers
The original CausalRivers paper builds a global causal graph using time-lagged causal relationships, but provides no specific time-lagged ground truth. It is a wrong statement, so its data cannot directly evaluate time-lagged causal graph recovery accuracy. We correct this statement.

Q8 Negative Edges
We use antisymmetrization to match the binary directed-graph evaluation of causalriver benchmark, which use 0/1 for presence of a directed causal relation. But it has limitation as you worried. For negative causal effects, antisymmetrization may change final edge direction. The same operation is applied to VAR and other baselines, including Appendix H.11, for consistent post-processing.

Q9 Experimental Settings
All hyperparameters were tuned on training/validation sets; no test set was used for selection. Validation split, please see reply to Reviewer fX4S Q4.

Q10 More RCA Baselines
Added more RCA baselines on page 9 and Appendix H.7. And LoCo remains competitive performance.

Q11 Creditcard & ε-Diagnosis
Creditcard train/test sample counts had a typo. In experiments we split data chronologically into first/second half with no overlap. ε-Diagnosis MSDS AC@3 also had a recording error. We reran and corrected.

Concerning about Weakness 7, the phenomenon is caused by the dataset itself. Creditcard anomalies are only ~0.17%, so overall detection is largely affected by normal samples. Thus Top-k has little effect on ROC-AUC, similar in baselines. This may relate to class distribution and detection difficulty, not Top-k failure. In contrast, Top-k has an obvious impact on performance in other datasets and tasks, such as PSM in AD and MSDS in RCA (Appendix H.4, H.8,H.9)

Reviewer_AJfyOpenReview ↗

Response to Authors’ Rebuttal

I thank the authors for the detailed response and the revisions. My main concern remains: with the support recovery result withdrawn, the theory reduces to the gradient-residual identity. The response also states that removing the gradient mask brings no increase, so it is still not shown that the proposed mechanism contributes beyond a plain linear forecaster on the real benchmarks. Other concerns, such as the hyperparameter selection, the consistency of the SWaT case study numbers, and the RCA attribution being redefined during the rebuttal, are also not completely resolved. I therefore keep my original rating.

We thank the reviewer for the follow-up comments. We understand that several points in our previous response were not sufficiently clear. We provide further clarification below.

1. Contribution of the Gradient Mask (GHT, Gradient Hard Threshold)

We would like to clarify that the gradient mask is a key component of LoCo, rather than an optional implementation detail. The Top-k mask controls which parameter coordinates are updated during training and is designed to guide the learning of time-lagged dependencies.

In Appendix H.11, we compare LoCo with Top-k=100%, which removes the gradient mask, under the same experimental pipeline on synthetic time-lagged datasets. Removing the mask does not improve performance; instead, performance decreases. This provides direct evidence that the gradient mask contributes to LoCo's time-lagged RCA and causal discovery performance.

We had already examined Top-k sensitivity on real datasets. As Top-k decreases, performance also shows a decline to some extent in conventional tasks, including anomaly detection on PSM and RCA on MSDS (Appendix H.4, H.8, and H.9). Although these sensitivity experiments are not a direct mask-removal comparison (Topk=100%) on every datasets, they provide additional evidence that the Top-k mechanism affects LoCo's practical performance.

2. Hyperparameter Selection

We clarify that hyperparameters were selected using only the training and validation data, without using the test set. The validation set contains normal samples, and only prediction MSE was used as the main criterion for model selection. The selected settings were then fixed for test evaluation. We have further clarified the validation split and experimental protocol in the revised manuscript.

3. SWaT Case Study

We rechecked the case study and corrected the variable-name error in Fig. 4(c): P502 should be P501. For case #2303, FIT401 is the abnormal variable identified by the detector, not P501.

We further clarify the difference between these quantities. The system-level SPE aggregates prediction errors over all 51 variables, whereas Algorithm 2 selects only the variable with the largest deviation at the current time step as the starting point for root-cause backtracking. In this case, FIT401 has a standardized prediction error of approximately 30, corresponding to a squared error of about 900, which is already substantially higher than the average error per variable. Therefore, the system-level SPE and the error of the selected variable naturally have different magnitudes.

Moreover, the SWaT system can have multiple sensors attacked simultaneously. Algorithm 2 selects the most deviant variable (FIT401 in this case) as the starting point; this does not mean that other variables, such as FIT402, are normal or have no contribution to the system anomaly. Backtracking from another highly deviating variable could also lead to changes in the system-level SPE.

Finally, P501's influence on FIT401 is not limited to the single time step t−11t-11. Its historical inputs at multiple earlier time steps, such as t−12t-12 and t−13t-13, may also contribute to the prediction error. The reported attribution magnitudes across t−11t-11 to t−15t-15 sum to approximately 200 (57+37+36+36+34), indicating a non-negligible contribution across multiple lags. These values measure changes in the system-level SPE under separate counterfactual replacements, rather than a direct decomposition of FIT401's squared error.

Since SWaT does not provide complete ground-truth root causes and propagation paths, we use the original study and the system's process structure to interpret the P501–FIT401 relationship. We therefore describe P501 as a process-grounded root-cause candidate, rather than a confirmed ground-truth root cause.

4. RCA Attribution

We understand the concern that the attribution may appear to have been redefined during the rebuttal. We would like to clarify that the implementation computes the model-internal contribution of a candidate historical variable by replacing its historical inputs with reference values, rerunning the prediction for the entire system, and measuring the absolute change in system-level SPE:

Attr(p∗)=∣SPE(t)−SPE(p∗)(t)∣.\mathrm{Attr}(p^*) =\left|\mathrm{SPE}(t)-\mathrm{SPE}^{(p^*)}(t)\right|.

This measures the magnitude of the change in prediction error after the replacement, regardless of whether the SPE increases or decreases. It is a model-based predictive attribution, not a physical intervention effect. During the rebuttal, we clarified the mathematical description and aligned the paper's explanation with the actual implementation to avoid ambiguity. We have also revised the relevant descriptions.

We appreciate the reviewer for pointing out these issues. These clarifications help us better distinguish the evidence for the gradient-mask (Gradient Hard Threshold) mechanism, the scope of the real-data case study, and the meaning of the model-based attribution.

Reviewer_AJfyOpenReview ↗

Response to Authors’ Clarification

I thank the authors for the additional clarifications. However, I believe the paper in its current form needs substantial changes before it is ready for publication. The theory and the SWaT case study no longer support the main claims, and the benefit of the gradient mask on real data is still not clearly shown. I therefore keep my original rating.

Reviewer_fX4SOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes LoCo, a lightweight and interpretable framework for multivariate time-series anomaly detection, root cause analysis (RCA), and causal discovery. The key idea is Local Competitive Learning, which applies a row-wise Top-k gradient mask to a single-layer forecasting model, encouraging a sparse time-lagged dependency structure. The authors introduce both Linear and Kernel variants, using the learned weights for anomaly detection through prediction residuals, historical root cause tracing, and global causal graph construction.

The method is evaluated on nine anomaly detection datasets, two RCA datasets (SWaT and MSDS), and CausalRivers for causal discovery. The results demonstrate competitive detection accuracy, computational efficiency, and promising variable-level RCA performance. Case studies further illustrate the proposed temporal root cause tracing capability.

Critical strengths
  1. Simple and computationally efficient framework with competitive performance.

The proposed Top-k gradient competition is conceptually simple and can be incorporated into a lightweight forecasting model without requiring complex architectures. Despite its simplicity, LoCo achieves competitive anomaly detection performance across multiple benchmarks while maintaining relatively low computational costs. This accuracy-efficiency trade-off is particularly valuable for resource-constrained streaming applications.

  1. Unified framework for anomaly detection and interpretable diagnosis.

The attempt to integrate anomaly detection, root cause localization, and time-lagged dependency discovery within a single model is practically meaningful. In particular, identifying potential upstream sources at earlier timestamps, rather than only highlighting anomalous variables at the detection timestamp, is a useful direction for industrial fault diagnosis. The SWaT case study provides an intuitive illustration of this capability.

Critical weaknesses
  1. Insufficient quantitative validation of the claimed time-lagged root cause tracing capability.

One of the main contributions of this paper is the ability to identify upstream root causes at specific historical timestamps and trace anomaly propagation through a time-lagged graph. However, the quantitative RCA evaluation primarily relies on conventional metrics such as AC@K on SWaT and MSDS, which assess root cause variable identification rather than the accuracy of historical root cause timestamps or propagation paths.

Similarly, the causal discovery evaluation is performed on a compressed variable-level graph rather than the original time-lagged graph. While the case studies provide illustrative examples of historical root cause tracing, they do not establish its accuracy through systematic quantitative evaluation.

As a result, the experiments demonstrate competitive conventional RCA performance but provide limited direct evidence for the proposed time-lagged RCA capability, which is presented as a central contribution of the paper. This gap between the claimed contribution and its empirical validation substantially weakens the experimental support for the proposed framework.

  1. Insufficient justification for interpreting learned predictive dependencies as causal relationships.

The paper interprets the learned forecasting weights as a time-lagged causal matrix and uses them for causal discovery and causal-driven root cause analysis. However, the weights are learned by minimizing forecasting errors, and their magnitudes primarily reflect predictive relationships between historical inputs and current outputs. Such relationships do not necessarily correspond to direct causal effects, particularly in the presence of correlated variables, indirect dependencies, or unobserved confounders.

Furthermore, the proposed counterfactual attribution procedure replaces a historical input with a model-predicted reference value and measures the resulting change in prediction error. While this provides a form of model-based attribution, it does not necessarily represent a causal intervention on the underlying physical system. In particular, it is unclear why the reduction in SPE after such replacement should consistently indicate the importance of a genuine upstream root cause.

Although the paper introduces structural assumptions to motivate causal interpretation, the experiments do not directly establish that the learned dependencies correspond to actual causal relationships. Therefore, claims regarding causal discovery, causal strength, and causal root cause tracing appear stronger than what the presented evidence supports.

Questions for authors
  1. Quantitative evaluation of time-lagged RCA (Clarification).

The current RCA evaluation primarily measures root cause variable identification. How can the proposed model's ability to identify the actual historical root cause timestamps and anomaly propagation paths be quantitatively validated? Are ground-truth fault injection times or propagation delays available in the evaluated datasets? If so, can the authors clarify whether the predicted temporal root causes agree with these annotations?

  1. Causal interpretation of the learned relationships (Clarification).

What justifies interpreting the learned forecasting weights as causal relationships rather than predictive dependencies? In particular, how does the method distinguish direct causal relationships from indirect dependencies or correlations induced by unobserved variables? Can the authors clarify the conditions under which the learned time-lagged graph can be interpreted as a causal graph in the real-world datasets?

  1. Validity of counterfactual root cause attribution (Clarification).

The RCA procedure replaces a historical input with its model-predicted reference value and ranks root causes based on the resulting reduction in SPE. Why should this reduction reliably reflect causal importance? In particular, how does the method handle cases where replacing a genuinely influential historical input increases the prediction error rather than decreasing it? Also, how is the predicted reference value guaranteed to represent normal behavior when its own historical inputs may already contain anomalies?

  1. Selection and stability of Top-k connections (Clarification).

How is the Top-k parameter selected for each dataset? Is it determined using only normal-data validation losses, or are downstream AD/RCA metrics involved in hyperparameter tuning? Since the learned connections are subsequently interpreted as causal relationships, how stable are the resulting graphs across different Top-k values and random initializations?

3 comments

We thank the reviewer for the careful reading and helpful comments. Based on these suggestions, we have revised the theoretical analysis, experimental evaluation, and related statements. The changes are marked in yellow in the revised manuscript.

Q1 Quantitative Evaluation

Existing real datasets mainly provide root-cause labels at the current time, but not complete time-lagged propagation paths or their ground truth. Therefore, the real cases in Section 5.3 are mainly used for analysis together with the original paper and actual system structure.

To address this, we built 3-node and 8-node synthetic datasets based on the Logistic equation. We simulated anomaly propagation by anomaly injection, so the propagation paths and time delays are known. We further introduced TLRCA@K and F1_{path} to evaluate time-lagged RCA, and TLROC to evaluate the recovery of time-lagged causal graphs. Dataset construction and metrics are in Appendix E.4, G.3, and G.5; results are in Appendix H.11.1.

Results show that Top-k gradient masking maintains prediction performance while balancing time-lagged RCA and CD tasks well.

Q2 Causal Interpretation

We reorganized the theory section to explain the link between gradients and causality more rigorously. In Section 4.1, under assumptions such as stationary local SVAR and finite lags, we analyze how learned weights relate to predictive dependence and to Granger causality.

Specifically, we use the Gradient–Residual Identity to link gradients, prediction residuals, and historical inputs. This shows how gradients reflect the incremental predictive information provided by past variables. We emphasize that this is a predictive-dependence signal related to Granger causality; it does not directly prove true structural causality.

Appendix D further discusses the difference between Granger causality and structural causality, and the difference in incremental predictive information between direct parents and indirect variables under the relevant assumptions. We also clearly state that unobserved confounding remains a limitation and a future direction. To validate time-lagged relations, we added TL3/TL8 synthetic dataset experiments with known causal structure and propagation delays; see Appendix H.11.1.

Q3 Counterfactual Attribution

LoCo’s counterfactual attribution measures how the current prediction error changes when a historical input is replaced by a reference value predicted by the model. It reflects model-internal predictive contribution.

(1) Why can this change reflect causal importance? Candidate parents come from the time-lagged predictive relation learned by LoCo on normal data. Under the assumptions in Q2, these dependencies reflect incremental predictive information from historical variables. Therefore, replacing a candidate node and observing the prediction-error change estimates its model-internal contribution to the current anomaly prediction, and is used to rank root-cause candidates.

(2) What if SPE increases after replacement? The original manuscript mainly described attribution as a decrease in SPE. However, the implementation considers both SPE increases and decreases, and uses the absolute change in SPE before and after replacement. Thus, attribution does not require SPE to decrease after replacement; it measures the importance of a historical input by the size of the prediction-error change. We have clarified this.

(3) What if the reference value itself is affected by anomalies? Reference values are generated by a predictor trained only on normal data, but its historical inputs may still be affected by earlier anomalies. Therefore, LoCo checks whether a candidate parent itself has a significant prediction deviation. If it is still abnormal, LoCo traces further upstream along time-lagged dependencies to find earlier root-cause candidates. This helps distinguish intermediate nodes in anomaly propagation from earlier anomaly sources, but it cannot guarantee finding the true physical root cause in all cases. We have clarified this process and its limitation in the revised manuscript.

Q4 Tuning

For all tasks—AD, RCA, CD, time-lagged RCA, and CD—we tuned parameters using only the training and validation sets, not the test set. The validation set contains only normal data, and is split from the last 10% of training data. Parameter selection is mainly based on prediction MSE, which is more feasible than other baselines in real scenario

In addition, because gradient masking can affect early gradient competition during training, initialization may affect LoCo training. We therefore use a small uniform initialization range [−0.00001,0.00001][-0.00001, 0.00001], and add comparisons of different initialization methods on time-lagged tasks in Appendix H.11.3.

I appreciate the authors' detailed response and efforts to address the reviewers' concerns. Most of my original concerns have been adequately addressed.

However, I remain unconvinced that the additional quantitative evaluation sufficiently validates the proposed time-lagged RCA capability. The synthetic datasets are small, all anomalies originate from the same root variable, and F1_path evaluates only a single propagation path. These limitations raise concerns about the generalizability of the results.

I also carefully considered AJfy's comments and find the methodological and experimental concerns raised there valid.

Overall, while the revisions have improved the paper, I would like to maintain my rating.

We thank the reviewer for the careful assessment and for recognizing our efforts to address the previous concerns. We appreciate the remaining comments on the time-lagged RCA evaluation and would like to further clarify the motivation behind our experimental design.

1. Synthetic Datasets and Experimental Design

We use Logistic systems because they provide controlled nonlinear dynamical systems in which time-lagged dependencies and ground-truth propagation relationships can be explicitly specified. This allows us to quantitatively evaluate time-lagged RCA and causal discovery, which is difficult with real-world datasets that lack complete ground-truth propagation paths. Logistic systems have also been used in prior studies to construct synthetic dynamical systems. Given the limited revision time, we focused on establishing this controlled quantitative evaluation. We further considered two system scales, TL3 and TL8, with 3 and 8 variables, respectively, to examine LoCo under different system sizes.

2. Single Root Cause and Path-level Evaluation

We deliberately use a single known root variable for anomaly injection to establish a clear starting point and ground-truth propagation path. This controlled setting allows us to examine whether LoCo can trace the propagation from the original anomaly source through time-lagged dependencies, without introducing ambiguity from multiple simultaneous sources.

Importantly, F1_path evaluates the agreement between the entire identified propagation path and its ground truth, rather than evaluating an individual edge or time lag in isolation. It therefore jointly assesses whether LoCo identifies the relevant variables, propagation relationships, and time lags along the path. Together with TLRCA@K and TLROC, these metrics evaluate root-cause localization, path reconstruction, and time-lagged causal graph recovery from complementary perspectives.

When multiple root causes are injected simultaneously, their propagation paths may overlap, making it more difficult to establish clear path-level ground truth and attribute downstream effects to individual sources. We therefore first use the single-source setting to provide a controlled evaluation of LoCo's time-lagged propagation-tracing capability.

3. Methodological and Experimental Concerns Raised by Reviewer AJfy

We have also carefully considered AJfy's comments and provided further clarifications in our latest response to Reviewer AJfy. In particular, we addressed the contribution of the gradient mask, hyperparameter selection, the SWaT case study, and the RCA attribution calculation. We kindly refer the reviewer to that response for the corresponding details.

We appreciate the reviewer's constructive feedback. We believe the additional experiments provide a useful quantitative evaluation of LoCo's time-lagged RCA capability under these settings, complementing the evaluations on real-world datasets in the original manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.