Pith. sign in

REVIEW 4 major objections 5 minor 16 references

WinTSR: A Windowed Temporal Saliency Rescaling Method for Interpreting Time Series Deep Learning Models

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read WinTSR presents a post-hoc interpretation method that rescales each feature's saliency by a time-relevance score, and reports that it outperforms ten baselines across five model architectures and three real-world datasets.

desk verdict A genuinely useful framework and a simple, fast method, but the 'significantly outperforms' claim outruns the evidence: the faithfulness proxy is aligned with the method's own masking, and there are no significance tests. read the letter →

arxiv 2412.04532 v3 pith:LAO5PSNT submitted 2024-12-05 cs.LG cs.AI

Md. Khairul Islam , Judy Fox This is my paper · ORCID
classification cs.LGcs.AI
keywords timeseriesinterpretabilitylocalexplanationtemporalsaliencyrescalingperturbation-basedattributionmodel-agnosticforecastingtransformerfoundationmodelcomprehensivenessandsufficiency
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to establish that a simple two-stage masking rule improves local explanations of multivariate time series models: instead of scoring each feature in isolation, it first asks how much each past time step matters to the model and then rescales every feature score by that time relevance. WinTSR masks all features at a given lag, measures the change in model output, normalizes those changes across the look-back window, and multiplies each individual feature's masking score by the resulting time weight. This makes temporal dependencies explicit in the attribution, such as the fact that the last observed hour or the same hour one day earlier often carries most of the predictive signal. The paper evaluates on three real-world datasets (electricity, traffic, and ICU mortality), five architectures including a transformer and an LLM-based foundation model, and ten comparison methods, reporting the best average rank for WinTSR in every dataset. If correct, the method gives practitioners a cheap, model-agnostic way to read which sensors and which historical moments drive a forecast or a clinical prediction, without training a surrogate model.

What carries the argument

The central object is the importance matrix $\phi_t = \{\phi_{j,l,t}\}$, where $\phi_{j,l,t} = \Delta^{\mathrm{feature}}_{j,l,t} \times \Delta^{\mathrm{time}}_{l,t}$. Here $\Delta^{\mathrm{time}}_{l,t}$ is the normalized magnitude of the change in model output when all features at lag $l$ are replaced by baseline values, and $\Delta^{\mathrm{feature}}_{j,l,t}$ is the change when only feature $j$ at lag $l$ is replaced. Both deltas are $\ell^1$ distances between original and perturbed predictions, and the baseline values are drawn from a normal distribution. The multiplication is the rescaling mechanism: it says a feature is credited only when both the feature and the time step it occupies move the model. The paper contrasts this with TSR, which rescales using a second attribution method's output, and with WinIT, which perturbs whole sliding windows and subtracts; WinTSR's product form is what carries the claimed gains.

What would settle it

Generate synthetic data with a known causal lag structure—say output depends only on feature 1 at lag 3 and feature 2 at lag 24, with all other entries pure noise—train any of the five models, apply WinTSR, and check whether the two true (feature, lag) positions receive the top importance scores across many seeds; if an exhaustive single-feature ablation recovers the known positions while WinTSR does not, the temporal rescaling step is not doing the explanatory work claimed.

Watch

Extended reading notes

Core claim

WinTSR claims that local explanations of multivariate time series are more faithful when feature importance is multiplied by temporal importance. The algorithm first masks every feature at each lag $l$ in the look-back window, records the $\ell^1$ change in model output, and normalizes these changes across time to obtain a time-relevance score $\Delta^{\mathrm{time}}_{l,t}$. It then masks each individual feature $x_{j,l,t}$, records a feature-relevance $\Delta^{\mathrm{feature}}_{j,l,t}$, and outputs $\phi_{j,l,t} = \Delta^{\mathrm{feature}}_{j,l,t} \times \Delta^{\mathrm{time}}_{l,t}$. The claimed discovery is that this rescaling captures delayed impacts between input and output better than either feature-only or time-only attributions, with runtime close to plain feature ablation and 32 to 367 times faster than TSR. The reported evidence is the best average rank in every dataset: $1.4 \pm 0.5$ on Electricity, $1.4 \pm 0.5$ on Traffic, and $2.4 \pm 1.5$ on MIMIC-III, on the comprehensiveness and sufficiency metrics.

Load-bearing premise

The load-bearing premise is that faithfulness is correctly measured by masking the top-ranked features (comprehensiveness) or keeping only them (sufficiency) and observing the output change; because WinTSR's scores are produced by exactly the same masking operation used in that evaluation, part of its reported advantage may reflect self-consistency rather than an independently verified explanation.

Editorial extensions

If this is right

  • Practitioners can obtain faithful local explanations for modern architectures, including transformer and LLM-based forecasters, without training a surrogate model.
  • The reported runtime makes temporal rescaling practical where TSR was not run at all, such as on the CALF foundation model.
  • Because the method is model-agnostic and output-based, the same code path works for multi-horizon regression and for binary clinical classification.
  • The open-source framework allows side-by-side benchmarking of 20+ time series models and 10+ interpretation methods, which the paper argues earlier evaluations lacked by relying on older LSTM/GRU baselines.
  • The time-relevance weights themselves are inspectable and show interpretable structure, with the highest weights at the most recent hour and the same hour one day earlier in electricity and traffic.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: the benchmark may reward methods whose scoring operation matches the masking used in evaluation, so a synthetic ground-truth test with known lags would be the cleanest way to confirm the claimed faithfulness.
  • Editorial extension: the rescaling step is separable from the underlying saliency estimator; applying the same time-weight product to gradient-based scores like integrated gradients could make those methods temporally aware without the cost of full TSR.
  • Editorial extension: the method returns a full (output, feature, lag) importance tensor, so it could support downstream uses the paper does not develop, such as grouping patients by which temporal signatures drive mortality or monitoring when a model's time-relevance curve shifts.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes WinTSR, a local post-hoc, model-agnostic interpretation method for multivariate time series models. WinTSR computes a time-relevance score for each look-back step by masking all features at that step and measuring the output change, then computes per-feature relevance by masking each feature individually and multiplies the two quantities to form the final importance score. The method is benchmarked on three real-world datasets (Electricity, Traffic, MIMIC-III) across five model architectures (DLinear, MICN, SegRNN, iTransformer, CALF) and compared with ten existing interpretation methods, using comprehensiveness and sufficiency derived from masking the top-k% features. The authors report average ranks showing WinTSR first on all three datasets and claim that it 'significantly outperforms' other local interpretation methods, while also releasing an open-source framework.

Significance. If the central claim is validated, WinTSR would be a useful, cheap, and model-agnostic interpretation tool for modern time series forecasting and classification models, and the companion open-source framework is a clear contribution to reproducibility and ease of benchmarking. The experimental scope, covering five architectures including a foundation model, is substantially wider than most prior work in this area. However, the headline claim rests on an evaluation protocol that is structurally aligned with the method's own definition, and the absence of statistical significance testing and of a ground-truth synthetic benchmark leaves the central claim unsupported. The paper's value as a framework and a careful empirical survey is real, but the evidence for 'significantly outperforms' is not yet convincing.

major comments (4)
  1. [5.1, Table 4] The abstract and Section 5.1 claim that WinTSR 'significantly outperforms' other methods, but no statistical significance testing is provided. Table 4 reports AOPCR/AOPC values and average ranks with standard deviations, but there are no confidence intervals on the AOPC/AOPCR values, no error bars over the three iterations mentioned in Table 8, and no paired test (e.g., Wilcoxon signed-rank) across models or datasets. The reported standard deviations on ranks reflect variation across the five models, not the sampling uncertainty of the comparison. Please add significance tests and report variance across seeds or bootstrap resamples.
  2. [3.2, Algorithm 1 vs 4.4] The evaluation metrics in Section 4.4 (comprehensiveness and sufficiency) are computed by masking the top-k% features and measuring the change in model output. WinTSR's importance scores in Algorithm 1 are defined by the same masking-and-output-change operation: the feature relevance is |f(Xt) − f(Xt \ x_{j,l})|. Thus, high comprehensiveness/sufficiency scores may partly reflect self-consistency between the method's definition and the evaluation metric, rather than external explanatory fidelity. Gradient-based and learned-mask methods are not constructed from this exact masking operation and are systematically disadvantaged. To support the claim, the authors should include a synthetic experiment with known ground-truth importance (e.g., a linear model or a controlled nonlinear generative process) and show that WinTSR recovers the true importance better than the baselines.
  3. [3.2, Eq. (2) and Algorithm 1] The product form of the importance score distorts feature ranking even in simple models. For a linear model f = Σ_{j,l} w_{j,l} x_{j,l}, the time-relevance score is Δtime_l = |Σ_j w_{j,l}(x_{j,l} − b_{j,l})|, so the final score φ_{j,l} = |w_{j,l}(x_{j,l} − b_{j,l})| × Δtime_l. This double-counts the aggregate effect of a time step, so a feature with small weight at a strongly perturbed time step can outrank a feature with larger weight at a quiet time step, contradicting the model's actual linear dependence. The paper does not include a synthetic linear benchmark to test whether this rescaling improves or distorts the recovered importance ranking. Please provide such a controlled experiment.
  4. [5.1, Table 4] The ranking procedure is not fully specified. The text says methods are ranked 'for each dataset and model in terms of overall comprehensiveness and Sufficiency' and then ranks are averaged, but it does not state whether comprehensiveness rank and sufficiency rank are averaged equally, how ties are broken, or whether the rank is computed per model before averaging. The conclusion is sensitive to this aggregation: for MIMIC-III, TSR has much higher comprehensiveness (0.80–0.90 vs WinTSR's 0.50–0.56) yet ranks 5th because of poor sufficiency. Please define the ranking rule explicitly and also report per-metric ranks separately.
minor comments (5)
  1. [Table 5] The model type labels in Table 5 are inconsistent with Section 4.2: MICN is the CNN-based model and SegRNN is the RNN-based model, but the table lists 'MICN (RNN)' and 'SegRNN (CNN)'. Please correct the labels.
  2. [Algorithm 1, Section 3.2] The 'featuregenerator()' baseline is not described precisely. The text says random values are drawn from a normal distribution, but the distribution parameters and whether the baseline is generated per instance or fixed per run should be stated. A sensitivity analysis of the baseline distribution would also strengthen the method's robustness claim.
  3. [5.4, Table 6] The MIMIC-III row in Table 6 is labeled 'Lookback = 36hr' while the column headers say 24hr and 48hr, and the text states that the lookback was varied to 24 and 36 hours for MIMIC-III. Please align the table header with the text.
  4. [Table 2 and Section A] The Electricity and Traffic datasets are listed as having five features, but these are one target series plus four time-encoded features (month, day, hour, day of week). This should be clarified in the main text, since the regression benchmark thus has very low feature dimensionality, which limits the generality of the conclusions for high-dimensional multivariate inputs.
  5. [Appendix B, Table 8] The training parameters report 'iterations 3' but the paper does not state how the three iterations are used (e.g., multiple random seeds, multiple runs, or multiple train/test splits). Please clarify and report the resulting variance in Table 4.

Circularity Check

1 steps flagged · score 5.0 of 10

WinTSR's leading rank is substantially self-consistency: the importance score and the faithfulness metric are the same masked-output-change operation, so the benchmark rewards the method's construction.

  1. self definitional [Section 3.1, Eq. (2); Algorithm 1; Section 4.4, Eqs. (3)-(5)]
    "ϕj,l,t = distance score(f (Xt), f (Xt \ xj,l,t)) (2) ... Algorithm 1: ϕj,l,t = ∆feature j,l,t × ∆time l,t ... Comp. = evaluation metric(f (Xi), f(Xi, \1:k)) (3)"

    WinTSR defines each feature's importance as the change in model output when that feature is masked. The evaluation defines faithfulness (comprehensiveness/sufficiency) as the change in model output when the top-scored features are masked, and the final claim is the rank on this metric. Thus the quantity WinTSR is built to maximize and the quantity used to judge it are the same perturbation operation; a method constructed from masked-output deltas is rewarded by construction. Gradient and Shapley-based methods, which optimize a different objective, are systematically disadvantaged. No synthetic ground-truth experiment is provided to show that masked-output change equals true feature importance, so the 'significantly outperforms' evidence rests on this self-consistent protocol.

full rationale

The paper is otherwise self-contained: the models are standard external benchmarks, the datasets are public (except MIMIC-III, which is a conventional private benchmark), and there is no load-bearing self-citation chain. The only substantive circularity is in the interpretation-evaluation loop: Eq. (2) defines importance as masked-output distance, and Eq. (3) measures faithfulness as masked-output distance after ranking by those scores. Since no ground-truth explanation is available for the three real datasets, the central claim that WinTSR 'significantly outperforms' rests on a metric that is aligned with the method's own construction. The comparison is still informative as a relative consistency check among methods, and Feature Ablation's second-place rank shows the effect is not unique to WinTSR, but the headline claim should be read as self-consistency under perturbation, not validated explanatory truth.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

WinTSR introduces no new entities, forces, or fitted parameters. It rests on two domain assumptions: perturbation-based output change represents feature importance, and the adopted faithfulness metrics are valid proxies for interpretation quality. The only hand-chosen design decision is the normal-distribution masking baseline. This is a modest axiom load for an empirical method paper.

free parameters (1)
  • Mask baseline distribution = standard normal
    Section 3.2 says random values from a normal distribution are used as the masking baseline because input features are normalized. This hand-chosen baseline affects all computed importance values and is not varied in the experiments.
assumptions (3)
  • domain assumption Output change after masking a feature indicates that feature's importance to the model.
    Eq. 2 defines importance as the L1 distance between original and perturbed predictions. This assumes the perturbation is meaningful and the model's reaction reflects genuine reliance on the masked input.
  • domain assumption Comprehensiveness and sufficiency, aggregated as AOPC/AOPCR, faithfully rank interpretation methods.
    Section 4.4 adopts these metrics with k in {5,7.5,10,15} because no ground-truth explanations exist for the real datasets. If the metrics favor certain perturbation behavior over true explanation quality, the headline comparison is weakened.
  • domain assumption Inputs are normalized as stated, so normal-distribution baselines are in-distribution perturbations.
    The validity of randomly generated normal baselines depends on the claim in Section 3.2 that input features are already normalized during preprocessing.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WinTSR: A Windowed Temporal Saliency Rescaling Method for Interpreting Time Series Deep Learning Models." pith.science (2026). https://pith.science/paper/LAO5PSNT

@misc{pith2026241204532,
  author       = {Pith},
  title        = {Pith review of: WinTSR: A Windowed Temporal Saliency Rescaling Method for Interpreting Time Series Deep Learning Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LAO5PSNT}},
  note         = {Machine review of arXiv:2412.04532}
}
read the original abstract

Interpreting complex time series forecasting models is challenging due to the temporal dependencies between time steps and the dynamic relevance of input features over time. Existing interpretation methods are limited by focusing mostly on classification tasks, evaluating using custom baseline models instead of the latest time series models, using simple synthetic datasets, and requiring training another model. We introduce a novel interpretation method, \textit{Windowed Temporal Saliency Rescaling (WinTSR)} addressing these limitations. WinTSR explicitly captures temporal dependencies among the past time steps and efficiently scales the feature importance with this time importance. We benchmark WinTSR against 10 recent interpretation techniques with 5 state-of-the-art deep-learning models of different architectures, including a time series foundation model. We use 3 real-world datasets for both time-series classification and regression. Our comprehensive analysis shows that WinTSR significantly outperforms other local interpretation methods in overall performance. Finally, we provide a novel, open-source framework to interpret the latest time series transformers and foundation models.

Figures

Figures reproduced from arXiv: 2412.04532 by the authors.

Figure 1
Figure 1. An overview of our framework with the proposed [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An illustration of (a) our proposed WinTSR [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Interpreted relevance scores (when τ = 1 or o = 1) for an arbitrarily chosen example from each dataset, some selected methods, and the iTransformer model. For MIMIC-III the overall most important features of each method are annotated. Both electricity and traffic features show high importance for more recent features and a 24-hour pattern for most methods. comprehensiveness score (DeYoung et al. 2019) for the classi… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 13 canonical work pages

  1. [1]

    (Suresh et al

    F eature Ablation (F A): The difference in output after replacing each feature with a baseline. (Suresh et al. 2017)

    work page 2017

  2. [2]

    (2020) ablated the input features by sampling counterfactuals from the bootstrapped distribution

    Augmented F eature Occlusion (AFO): Tonekaboni et al. (2020) ablated the input features by sampling counterfactuals from the bootstrapped distribution. Table 9: Experiment configuration of the models. Model Hyper-parameters Parameter Value Parameter Value DLinear moving avg 25 MICN encoders 2 kernel (18, 12) decoders 1 dimension 128 SegRNN layers 1 dimens...

    work page 2020

  3. [3]

    F eature Permutation (FP): Permutes the input feature values within a batch and computes the difference between original and shuffled outputs (Molnar 2020)

    work page 2020

  4. [4]

    Integrated Gradients (IG): Sundararajan, Taly, and Yan (2017) assigned an importance score to each input feature by approximating the integral of gradients of the model’s output to the inputs

    work page 2017

  5. [5]

    Gradient Shap (GS): Lundberg and Lee (2017) approximated SHAP values by computing the expectations of gradients by randomly sampling from the distribution of baselines/references

    work page 2017

  6. [6]

    Dyna Mask (DM): Crabbe and van der Schaar (2021) learned masks representing feature importance

    work page 2021

  7. [7]

    Extremal Mask (EM): Enguehard (2023a) improved the static perturbation from Dyna Mask by learning not only masks but also associated perturbations

    work page 2023

  8. [8]

    (2023) explicitly accounted for the temporal dependence among observations of the same feature by summarizing its importance over a lookback window

    Windowed F eature Importance in Time (WinIT):Leung et al. (2023) explicitly accounted for the temporal dependence among observations of the same feature by summarizing its importance over a lookback window. Table 11: Runtime (minutes): The complete comparison of the interpretation methods averaged across iterations. WinTSR has a runtime similar to other p...

    work page 2023

Show all 16 references
  1. [9]

    (2020) proposed to separate the temporal dimension when calculating feature importance and rescaling it

    Temporal Saliency Rescaling (TSR): Ismail et al. (2020) proposed to separate the temporal dimension when calculating feature importance and rescaling it

    2020

  2. [10]

    (2024d) designed a contrastive learning-based masking method to learn locally sparse perturbations for better explaining feature relevance with and without top important features

    ContraLSP: Liu et al. (2024d) designed a contrastive learning-based masking method to learn locally sparse perturbations for better explaining feature relevance with and without top important features. D Time Complexity Table 11 shows the full run time comparison between the i...

    2024

  3. [14]

    CALF: Aligns LLMs for time series forecasting with cross-modal fine-tuning (Liu et al. 2024a)

  4. [15]

    TimeLLM: Reprograms LLMs and its tokenization for better forecasting (Jin et al. 2024)

    2024

  5. [16]

    GPT4TS: Generalizes pretrained LLMs (GPT-2, Bert) for time series (Zhou et al. 2023). We include the following transformer-based and other recent time series models in our proposed framework: (1) Transformer, Vaswani et al. (2017) (2)Reformer, Kitaev, Kaiser, and Levskaya (202...

    2017

  6. [2021]

    International Journal of Forecasting, 37(4): 1748–1764

    Temporal fusion transformers for interpretable multi-horizon time series forecasting. International Journal of Forecasting, 37(4): 1748–1764. Lin, S.; Lin, W.; Wu, W.; Zhao, F.; Mo, R.; and Zhang, H. 2023. Segrnn: Segment recurrent neural network for long-term time series fore...

    2023 arXiv

  7. [2022]

    arXiv preprint arXiv:2202.01381

    Etsformer: Exponential smoothing transformers for time-series forecasting. arXiv preprint arXiv:2202.01381. Wu, H.; Hu, T.; Liu, Y .; Zhou, H.; Wang, J.; and Long, M. 2022. Timesnet: Temporal 2d-variation modeling for general time series analysis. arXiv preprint arXiv:2210.021...

    2022 arXiv

  8. [2023]

    Nature Machine Intelligence , 5(3): 250–260

    Evaluation of post-hoc interpretability methods in time-series classification. Nature Machine Intelligence , 5(3): 250–260. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, Ł.; and Polosukhin, I. 2017. Attention is all you need. Advances in...

    2017 arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.