acceptodds
Under review as a conference paper at ICLR 2027

SLRCA: Lightweight Root Cause Analysis of Multivariate Time Series via Parameter Sharing

Abstract

Tracing the root causes of anomalies in multivariate time series is an important and challenging task. Root cause localization requires modeling each variable's normal behavior, but assigning a separate neural network to each variable duplicates temporal prediction parameters, leading to overfitting and parameter proliferation. We propose SLRCA (Shared Lightweight Root Cause Analysis), a lightweight framework for root cause localization through parameter sharing. SLRCA uses vector autoregression (VAR) to model variable-specific linear lagged dependencies, then uses a single temporal convolutional network (TCN) to refine the linear predictions from each variable's residual history. Together, the two components predict each variable's normal behavior; the model uses deviations of actual observations from these predictions to rank candidate root causes. Across four types of synthetic systems and three real-world datasets, SLRCA achieves excellent performance while using only approximately 0.1%–1% as many model parameters as the million-parameter comparison methods.

Primary area:
learning on time series and dynamical systems
Keywords:
root cause analysis, multivariate time series, parameter sharing, anomaly diagnosis, temporal convolutional networks

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_JF6SOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes SLRCA, a lightweight method for root cause localization in multivariate time series. A ridge-fitted VAR(1) captures the linear lagged dependencies with variable-specific coefficients, and a single small TCN shared across all variables predicts the remaining VAR residual from each variable's own residual history and three cross-variable summary statistics. Variables are then ranked by their normalized prediction errors over the fault window, using a persistent-shift score combined with a pulse score calibrated on normal data. A proposition bounds the risk gap between shared and variable-aware predictors, and a toy example shows how VAR residualization can make sharing lossless. Experiments on four synthetic systems and SWaT, MSDS and WADI, with seven deep baselines and several ablations, report AC@1 of 0.667 on SWaT and 0.652 on MSDS using roughly 9k to 25k parameters in total.

Critical strengths
  • The method is simple and very compact (around 9k to 25k parameters in total), which is relevant for deployment on shared or limited hardware.

  • On SWaT and MSDS, the architecture-controlled comparisons in Table 5 support the design. VAR + shared TCN beats VAR alone, raw + shared, VAR + independent TCNs and the PatchTST variant under the same scoring.

  • Testing the fixed models on newly generated fault trajectories, without refitting or recalibration, is a useful check that the results are not tied to the original event set.

  • Proposition 1 appears correct under its assumptions, and the toy example in A.3 gives clear intuition for why residualizing before sharing can remove conflicting prediction targets.

Critical weaknesses
  • The combined architecture mainly helps on SWaT and MSDS, only 43 events in total. VAR alone matches or nearly matches it on three synthetic datasets, Raw+Shared is better on Lorenz96, and VAR alone has higher MRR on WADI (Table 5). Hence, I am not convinced the improvement generalizes beyond these two datasets.

  • Only 20 SWaT attacks are evaluated, without clear selection criteria. The dataset providers list 41 attacks, of which 36 are considered valid and are used in other RCA works on SWaT. AERCA uses the same count of 20, but are these the same attacks and labels? Please list the selected attacks, explain the exclusions, and report results on all 36 valid attacks.

  • The MSDS results are unusual: eight configurations (five baselines in Table 2, three variants in Table 5) score exactly zero AC@1 with only ten candidates, while SLRCA gets 15/23 correct with zero variance across seeds. Please provide the label distribution, a constant-ranking baseline, and per-event top-ranked variables.

  • How were the dataset-specific settings in Table 3 selected? Normal validation loss does not explain the scoring-window choice. The MSDS drop from 0.652 to 0.507 ± 0.153 under the alternative settings in Table 6 makes this important. Were fault events used for tuning?

  • Scoring appears to account for a substantial part of the gains. VAR alone with score S beats every published baseline in MRR on SWaT and WADI. What happens when compatible baselines receive the same calibration and scoring?

  • Baseline training needs justification: SARAD gets up to three epochs, TranAD five, and PatchTST-RCA 400 to 800 updates. Were these models trained to convergence? AERCA is missing on WADI without explanation, and lightweight baselines such as n-sigma, RCD, and CIRCA, as well as recent RCA methods evaluated on SWaT such as ShaTS[1] and CondAttr[2], are missing.

  • Using faulty history may allow persistent faults to become predictable, weakening the residual at the source. Does evaluation start at the true fault onset? The Lorenz96 failure makes diagnosis-delay and window-length sensitivity important.

  • Important design choices lack ablations. Does spectral scaling to 0.9 hurt slowly varying industrial signals? Do the shared summary channels spread fault effects into other variables' predictions? Constant-channel handling and actuator root causes also need clarification.

  • Proposition 1 does not establish that fitted VAR reduces the sharing gap or that sharing improves root-cause rankings, so it offers limited support for the proposed architecture. The novelty is also unclear given prior work on neural residual prediction and per-series statistical models with shared neural networks [3], neither of which is discussed.

References:

[1] Franco, M., Gómez, Á. L., & Maimó, L. F. (2025). ShaTS: A Shapley-based Explainability Method for Time Series Artificial Intelligence Models applied to Anomaly Detection in Industrial Internet of Things. ArXiv. https://arxiv.org/abs/2506.01450

[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] Smyl, Slawek. "A hybrid method of exponential smoothing and recurrent neural networks for time series forecasting." International journal of forecasting 36.1 (2020): 75-85.

Questions for authors
  • Which 20 SWaT attacks are used, and why were the others excluded? How does SLRCA perform on all eligible attacks, with any exclusions explained?

  • What is the root-cause label distribution on MSDS? How does a constant-ranking baseline perform, and which variable does each method rank first in each event?

  • How were the settings in Table 3 selected, in particular the scoring-window length, clipping and scale estimator? Were fault events used at any stage?

  • If compatible baselines such as GDN or TranAD use the same normalization and score S, how much of the margin remains?

  • Were SARAD, TranAD and PatchTST-RCA trained to convergence? Why is AERCA missing on WADI?

  • Does the scoring window start at the true fault onset? How do the results change if diagnosis starts a few steps later?

  • The full model offers little benefit over VAR on three synthetic datasets and WADI, while Raw+Shared is stronger on Lorenz96. Is there a measurable property of the data that tells in advance whether the shared TCN will help?

  • Is there any evidence that the VAR residuals on the real datasets come closer to the common-optimum condition of Proposition 1 than raw observations? Shared versus per-variable validation loss would be indirect evidence, but even that is not reported.

  • Nonlinear cross-variable effects can only enter the shared TCN through permutation-invariant summaries, which lose which variable produced them. How would the method capture nonlinear interactions that VAR cannot, and is there evidence that it does so on the real datasets?

  • SWaT and WADI are closed-loop systems where a source and its controller response can deviate in the same step. What assumptions does the method need to separate them, and what is the intended scope of the RCA claim?

Additional suggestions

  • Ablations without the three summary channels and without spectral scaling on the real datasets.

  • Per-variable calibration of the persistent score on normal windows, since residual autocorrelation can differ across variables.

  • Runtime and memory comparison with the baselines, not only for SLRCA.

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

This paper introduces SLRCA, a lightweight framework for root cause localization in multivariate time series in resource-constrained industrial environments. SLRCA employs a two-part architecture: a variable-specific VAR to capture linear lag dependencies, and a single shared TCN to predict nonlinear structures in the VAR residuals. Root causes are ranked using a combination of persistent-shift and pulse scores. Results show that SLRCA achieves competitive performance on most datasets while using far fewer parameters.

Critical strengths
  1. This paper focuses on resource-constrained environments, which is practical for industrial settings. It substantially reduces the number of model parameters while achieving competitive performance according to experimental results.
  2. The architecture design is intuitive and easy to implement. It uses variable-specific VAR for linear lag dependencies while using a TCN for shared VAR residuals. The parameter-sharing strategy naturally saves parameters for the whole network.
  3. The experiments on seven datasets, including both synthetic and real-world benchmarks, show competitive results for SLRCA.
Critical weaknesses
  1. The novelty of the proposed method is relatively limited. SLRCA appears to be a sound engineering combination: variable-specific VAR and a shared TCN. The concept of parameter sharing has been extensively explored in time series modeling, despite the paper's claim of being the 'first to combine variable-specific linear prediction with parameter sharing for RCA'.

  2. Although VAR can be viewed as a linear causal structure, the paper does not deeply explore its causal semantics. SLRCA's root cause scoring is primarily based on the statistical deviation of prediction residuals, rather than modeling 'anomaly propagation paths' or 'shifts in causal relationships.' The method may lack interpretability in complex propagation scenarios, especially in industrial environments.

  3. The performance advantages are inconsistent, and the method is not optimal on certain datasets. For example, SLRCA is not consistently leading on some metrics (e.g., AC@3, AC@5, MRR). On WADI, SLRCA ties with PatchTST-RCA on AC@1 but has a slightly lower MRR. Some synthetic datasets have near-perfect scores, limiting the ability to differentiate method performance.

  4. Concerns regarding experimental fairness and reproducibility details. Given the limited number of events in the real-world datasets (e.g., only 11 fault events in WADI), the statistical stability of the results is questionable. This is particularly concerning since the paper emphasizes its focus on real-world industrial and resource-constrained scenarios, yet the evaluation relies on only three real-world datasets. Furthermore, Appendix B.2 / Table 3 reveals a significant disparity in training protocols: SARAD and TranAD are trained for at most 3 and 5 epochs, respectively, whereas the TCN in SLRCA is trained for up to 150 (MSDS) and 2000 (SWaT) epochs. Such a vast difference in training may raise concerns about the fairness of the comparison.

  5. The presentation of the paper needs improvement. For example, in lines 183–198, the authors write: 'A shared network can produce different predictions from different histories. We now consider two variables with identical historical inputs but different prediction targets.' Does the example aim to illustrate the effect of identical or different histories? In addition, the choice of √2 in the example is not explained in the main text and appears arbitrary on first reading, only the appendix makes clear that this value is part of a standardized toy construction. As written, this example feels abrupt and is difficult to follow without consulting the appendix.

Questions for authors

See Weaknesses.

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

The paper proposes SLRCA, a compact framework for root cause localization in multivariate time series. A variable-specific VAR model captures linear lagged dependencies, and a shared temporal convolutional network predicts the remaining residual dynamics. Candidate variables are ranked using calibrated scores for persistent shifts and short pulses. The paper also characterizes when a shared predictor can attain the same population risk as an identity-aware predictor under Smooth-L1 loss. Experiments on four synthetic systems and three real-world datasets demonstrate favorable accuracy–parameter tradeoffs, with the strongest evidence for the combined architecture on SWaT and MSDS.

Critical strengths
  1. A clear and economical modeling design. Separating variable-specific linear coefficients from shared temporal prediction is well motivated and easy to understand.

  2. Informative ablations beyond comparisons with large baselines. Table 5 includes VAR-only, raw-input and residual-input models, shared and independent networks, and variable-identity embeddings. Table 10 isolates the scoring contribution while holding predictors fixed. These controls support complementarity between VAR and shared prediction on SWaT and MSDS. The paper also reports counterexamples, including stronger raw-input sharing on Lorenz96 and no AC@1 improvement over VAR alone on WADI.

Critical weaknesses
  1. The evidence supporting the VAR–sharing decomposition is narrower than its methodological motivation. Proposition 1 fixes the residual representation and compares unrestricted measurable predictors; it does not establish that fitted VAR generally reduces prediction conflicts or explain the finite-sample behavior of the shared TCN. Empirical benefits are also system-dependent: Raw+Shared outperforms SLRCA on Lorenz96, while VAR alone matches its AC@1 and achieves higher MRR on WADI. The paper should more clearly delimit when the decomposition helps and distinguish its conditional theoretical motivation from the demonstrated empirical gains.

  2. The adequacy of baseline training is insufficiently established. Table 4 limits SARAD and TranAD to three and five epochs, respectively, whereas SLRCA allows up to 2,000 epochs on SWaT. Different epoch counts do not inherently imply unfairness, but the manuscript lacks convergence curves, selected-checkpoint information, and sufficiently documented tuning procedures to establish that the decisive baselines were adequately optimized. This limits confidence in the reported performance advantages over published methods.

Questions for authors
  1. Scope of the VAR–sharing decomposition: What characteristics of the evaluated systems explain when VAR residualization benefits shared prediction, particularly its gains on SWaT/MSDS versus the Lorenz96/WADI results? Can the existing experiments substantiate this explanation, while distinguishing the fixed-representation population result in Proposition 1 from the behavior of the fitted VAR and finite shared TCN?

  2. Baseline training adequacy: Could you provide the configuration-selection procedures, normal-validation learning curves, and selected checkpoints for the baselines driving the main comparisons, particularly SARAD and TranAD? For runs still improving at the training cap, does extending training materially change the reported performance gaps?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.