TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation

Changhun Kim; Eunho Yang; Hyeongwon Jang

REVIEW 4 major objections 6 minor 2 cited by

TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper claims that time-series explanations should carry a sign for each point, that standard evaluation metrics hide the value of signed attributions by removing points simultaneously and letting positive and negative contributions…

desk verdict A real problem with a plausible fix, but the empirical claim stands on an unvalidated metric and a statistically marginal gain over IG. read the letter →

arxiv 2506.05035 v1 pith:ZSTGOBZW submitted 2025-06-05 cs.LG cs.AI

Hyeongwon Jang , Changhun Kim , Eunho Yang This is my paper

classification cs.LGcs.AI

keywords timeseriesexplainabilityIntegratedGradientsdirectionalattributionfaithfulnessmetricscumulativepredictiondifferencesegmentationmaskingMIMIC-IIImortalityexplainableAI

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's central claim is that time-series explanations should carry a sign—whether each time point pushes the prediction up or down—and that current evaluation metrics hide this information because they remove the top-K important points simultaneously, letting positive and negative contributions cancel. To fix that, the paper defines Cumulative Prediction Difference (CPD) and Cumulative Prediction Preservation (CPP), which remove points one at a time and accumulate the change in the predicted-class distribution; under these metrics, conventional Integrated Gradients (IG) already beats recent masking-based explainers. The paper then proposes TIMING, which replaces IG's zero baseline with randomly retained whole segments so that the integration path can break temporal relationships without leaving the data distribution, while preserving IG's sensitivity and implementation-invariance guarantees. On real-world benchmarks (MIMIC-III, PAM, Boiler, Epilepsy, Wafer, Freezer), TIMING attains the highest CPD scores, including 0.366 at K=50 and 0.505 at K=100 on MIMIC-III, and roughly matches IG on synthetic saliency benchmarks. If correct, this means segment-aware, directional attribution is the more faithful and practical way to explain time-series classifiers.

What carries the argument

The load-bearing object is the masked integrated-gradients path. For a binary mask $M\in\{0,1\}^{T\times D}$, TIMING's building block is $\mathrm{MaskingIG}_{t,d}(x,M)=x_{t,d}M_{t,d}\int_0^1 \partial F_{\hat y}(\alpha(M\odot x)+(1-M)\odot x)/\partial x_{t,d}\,d\alpha$, where $F_{\hat y}$ is the predicted-class probability. The intermediate point $\alpha(M\odot x)+(1-M)\odot x$ keeps some original values (where $M_{t,d}=0$) and scales others toward zero, so temporal relations can break while the input stays closer to the data distribution. The mask generator $G(n,s_{\min},s_{\max})$ samples $n$ contiguous segments of length between $s_{\min}$ and $s_{\max}$; the final attribution is the expectation over masks conditional on each point being scaled, i.e. masked. Proposition 4.1 lets this expectation be evaluated as a single path with one random mask per $\alpha$ step, which is what makes TIMING cost about the same as IG.

What would settle it

Take TIMING's MIMIC-III attributions, keep all magnitudes, shuffle or flip the signs, and recompute CPD at K=50; if the perturbed version scores close to 0.366, the metric rewards sign alternation rather than point localization, and the paper's central comparison collapses.

Watch

Extended reading notes

Core claim

The central claim is that signed, cumulative evaluation changes the ranking of time-series explainers: when you remove the highest-attribution points one at a time and sum the changes to the full predicted-class distribution (CPD), and likewise remove the lowest-attribution points (CPP), Integrated Gradients outperforms recent masking-based methods such as Extrmask and ContraLSP, because points with positive and negative influence no longer cancel. The paper then argues that plain IG is suboptimal for time series because its zero baseline scales all points together, keeping temporal relations intact and sending interpolated inputs out of distribution. TIMING replaces that baseline with randomly retained segments: a mask M keeps some whole segments at their original values while scaling the rest, and the attribution is the expected gradient path conditional on the point being masked, computed with one random mask per integration step. This keeps IG's sensitivity and implementation-invariance axioms, drops completeness, and yields the best CPD scores on MIMIC-III (0.366 at K=50, 0.505 at K=100), PAM, Boiler, Epilepsy, Wafer, and Freezer, while matching IG on synthetic saliency benchmarks.

Load-bearing premise

The entire ranking depends on CPD and CPP actually measuring explanation faithfulness, and the paper itself warns that signed attribution methods can inflate CPD by alternating positive and negative points in the removal order.

Editorial extensions

If this is right

Under CPD and CPP, recent results that rank unsigned masking methods such as Extrmask, ContraLSP, and TimeX++ ahead of gradient methods need to be re-examined, because their advantage was measured with simultaneous removal that lets opposing contributions cancel.
Practitioners can apply TIMING to any differentiable time-series classifier at roughly the cost of Integrated Gradients (about 0.04 seconds per sample in the paper's measurements) and obtain explanations that retain directional information.
Because TIMING deliberately gives up IG's completeness axiom, the sum of its attributions is not guaranteed to equal the prediction difference; users who rely on summed attributions for feature budgets should prefer plain IG.
TIMING's largest gains appear under zero substitution, where masked points become zero, indicating that it captures points whose removal genuinely changes the model, while remaining competitive under average substitution.

Reading between the lines

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

The paper's own warning about signed methods inflating CPD by alternating positive and negative removal order implies a control experiment it does not run: report CPD for TIMING, IG, and GradSHAP with signs randomly flipped or with removal order shuffled; without that control, the CPD leaderboard may partly reflect metric artifacts.
Segment-based masking is a generic structural prior, so the same recipe could improve Integrated Gradients for images (object-region masks) or graphs (node-community masks), where the zero baseline suffers from the same out-of-distribution and relation-preservation problems.
A direct clinical follow-up would test whether TIMING's lactate attributions on MIMIC-III change a clinician's trust or decision time compared with unsigned baselines; the paper shows qualitative agreement with clinical knowledge but does not include a human-subject study.
If CPD and CPP are accepted as faithfulness metrics, many existing time-series XAI comparisons are confounded, and re-benchmarking published methods with cumulative removal would be a low-cost, high-value re-analysis.

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.

Desk Editor's Note

read the letter

Two things to know about this paper. First, it has a real idea: the standard practice of removing top-K important points simultaneously can cancel out positive and negative contributions, making IG look bad. Second, the paper's own solution—CPD/CPP metrics plus a segment-masked IG variant (TIMING)—is plausible but the empirical case is conditional on accepting those metrics, and even there TIMING's edge over plain IG is within error bars.

What's genuinely good: the cancellation observation is correct and worth stating; CPD/CPP are a reasonable response, essentially summing per-step prediction changes instead of one net change. TIMING is a sensible adaptation of Expected Gradients to time series, using segment-level random masks to preserve temporal structure and mitigate OOD. The experiments are extensive, the code is released, and the hyperparameter sensitivity analysis is careful. The paper does a service by pointing out that signed gradient methods have been underrated.

The soft spots are real but not fatal. On the standard metrics (Acc, CE, Suff, Comp), TIMING is not the best; on synthetic benchmarks with known ground truth, it matches IG on AUP/AUR and trails ContraLSP/Extrmask. The authors explain this away as those methods sacrificing model faithfulness, but that rationale is only convincing if you already trust CPD. The advantage over IG on MIMIC-III (0.366 vs 0.342 at K=50) is about one standard error, with no significance testing. The theoretical propositions are near-tautological: Fubini interchange (Prop 4.1) and standard IG axioms (4.2-4.3), with the authors explicitly giving up completeness. The stress-test's specific mechanism for sign-alternation inflation doesn't fully hold up—the L1 norm per step doesn't depend on the sign of the change—but the paper itself warns about ordering manipulation and then doesn't control for it, so the concern is legitimate at the level of metric validity.

Bottom line: This is a solid, citable contribution to time series XAI evaluation, but the headline performance claim needs more work. The right referee would ask for external validation of CPD/CPP (e.g., whether they correctly rank methods on synthetic data with known importance), significance tests, and a direct test of sign-alternation effects. I'd send it to review.

Referee Report

4 major / 6 minor

Summary. The paper makes three contributions. First, it proposes two evaluation metrics, CPD (Cumulative Prediction Difference) and CPP (Cumulative Prediction Preservation), which sequentially remove the highest- and lowest-ranked attribution points and sum the ℓ1 distances between consecutive model output probability vectors. The motivation is that conventional metrics, which remove top-K points simultaneously, cancel out opposing positive and negative attributions and thus bias rankings toward methods whose attributions are aligned in sign. Second, the paper observes that standard Integrated Gradients (IG) performs well under CPD/CPP, and third, it introduces TIMING, a variant of IG that replaces the uniform zero baseline with segment-based stochastic masks, drawing attribution from an expectation over random segment masks. The paper proves that TIMING retains sensitivity and implementation invariance, notes that it drops completeness, and reports experiments on synthetic and real-world benchmarks, claiming that TIMING outperforms existing time-series XAI baselines.

Significance. If the central claims hold, the CPD/CPP metrics provide a new and arguably more faithful way to evaluate directional time-series attributions, and TIMING offers a lightweight, gradient-based alternative to recent learned perturbation methods. The paper is commendably transparent in places: it acknowledges that signed methods can inflate CPD through alternating signs, and it explicitly states that TIMING does not satisfy the completeness axiom. However, the manuscript's central empirical assertion—that TIMING outperforms existing baselines—rests on a metric that the paper itself identifies as gameable, and the quantitative differences over IG on the main benchmark are within one standard error. The theoretical claim that TIMING preserves IG's sensitivity is supported by a proof that appears to contain an error. Because the method's practical value depends on the validity of CPD/CPP and the correctness of the axiomatic argument, the result is plausible but not yet established at the standard required for acceptance.

major comments (4)

[Section 3.2, Table 2] The headline claim that TIMING outperforms existing time-series XAI baselines rests on CPD/CPP, but Section 3.2 warns that for signed methods, 'alternating positive and negative points by absolute value order can inflate the metric.' TIMING is a signed gradient-based method, while the main competitors (Extrmask, ContraLSP, TimeX++, Dynamask) produce unsigned attributions. The comparison therefore falls exactly in the regime the paper flags as vulnerable. The paper does not test whether TIMING's top-K attributions alternate in sign, does not provide a control (e.g., using absolute values for the removal order), and does not otherwise correct for this potential bias. In addition, the TIMING-versus-IG differences (CPD 0.366±0.021 vs 0.342±0.021 for K=50; 0.505±0.029 vs 0.469±0.030 for K=100) are within about one standard error, and no significance testing is reported. These issues together undermine the central empirical claim.
[Section 4.2, Proposition 4.2] The proof of Sensitivity is incorrect as written. The proof states that when x and x' differ only at coordinate (t,d), TIMING reduces to IG with baseline x' for every mask with Mt,d=1, and therefore inherits IG's sensitivity property. However, the definition of MaskingIG in Section 4.2 uses baseline (1−M)⊙x, not (1−M)⊙x + M⊙x'. For a generic alternative input x' (not equal to 0), the equality (1−M)⊙x = x' does not hold, and the subsequent identification of TIMING with IG is unjustified. The proposition may be true with a different argument, but the provided proof does not establish it.
[Section 5.2, Tables 2 and 4] The claim that TIMING 'outperforms existing time series XAI baselines' is broader than the evidence. In Table 2, on the standard metrics (Acc, CE, Suff, Comp), TIMING is not the best: Extrmask and ContraLSP score substantially better on Acc and Comp, and several baselines have better CE. On the synthetic datasets with known ground truth (Table 4), TIMING's AUP/AUR are approximately equal to IG and below ContraLSP and Extrmask. The paper responds that CPD is the more meaningful criterion, but CPD is introduced in the same paper, and no evidence is provided that CPD correlates with actual faithfulness (e.g., recovery of ground-truth salient points). Without an independent justification of the metric, using CPD as the sole arbiter for the central comparison is circular.
[Section 5.2, Tables 2 and 3] Even within CPD, the paper's claim of consistent state-of-the-art performance is overstated. On MIMIC-III the TIMING gain over IG is within one standard error, and on Freezer (Table 3) TIMING's CPD (0.409±0.109) is statistically indistinguishable from IG (0.405±0.111). The large relative gains reported (e.g., 109.8% for Boiler) are driven by low baseline values rather than large absolute differences, and no confidence intervals or significance tests are provided. The text should either qualify the claim to specific datasets and metrics or supply the missing statistical analysis.

minor comments (6)

[Figure 1] The plots in Figure 1 lack axis labels and a legend that clearly maps colors to methods; as presented, the reader cannot verify the claimed ordering between panels (c) and (d).
[Table 1] The rows 'IG (Unsigned)' and 'IG (Signed)' are not defined in the text; presumably the former is the absolute value of IG attributions, but this should be stated explicitly because the distinction is crucial to the argument.
[Algorithm 1] Line 13 reads 'grad s:c,d ← 0'; the subscripts appear to be a typo (likely 's:e,d' or the intended segment range). Line 17's division by 'n_samples − unmaskcount' should be explained, as the relationship to the expectation in Proposition 4.1 is not immediate.
[Sections 4.1–4.3] The notation for x' is inconsistent: in Section 4.1 and Proposition 4.1, x' is the baseline (later specialized to 0), while in Proposition 4.2, x' denotes an alternative input. This makes the proof of Proposition 4.2 difficult to follow and should be clarified.
[Section 5.2] The sentence 'TIMING consistently achieves state-of-the-art performance across all datasets' is contradicted by the Freezer row in Table 3, where TIMING and IG are within one standard error; the claim should be softened accordingly.
[Tables 5 and 6] The ablation and hyperparameter tables show that all conditions produce CPD values within one standard error of each other; the text interprets the small differences as substantive, but it should acknowledge that the differences are not statistically distinguishable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TIMING's derivation is self-contained, and the CPD metric concerns are evaluation-validity issues rather than definitional loops.

full rationale

TIMING is explicitly constructed from Integrated Gradients: TIMING_{t,d}(x;n,s_min,s_max) = E_{M~G}[MaskingIG_{t,d}(x,M) | M_{t,d}=1]. Proposition 4.1 is a Fubini-Tonelli interchange of an integral and a conditional expectation applied to that definition, so it does not assume the conclusion. Propositions 4.2-4.3 are proved by reducing TIMING to IG under the stated one-point masking condition, and Proposition 4.4 openly disclaims completeness, so the 'maintaining theoretical properties' claim is not hiding a circular assumption. CPD and CPP are defined for arbitrary attribution matrices and are not functions of TIMING's parameters; therefore, reporting TIMING's CPD score is not a definitional equivalence. The paper itself flags a real confound in Section 3.2: 'for signed methods, alternating positive and negative points by absolute value order can inflate the metric; they should avoid using attribution signs to manipulate point ordering.' TIMING is signed, so this is a genuine threat to the head-to-head CPD comparison against unsigned baselines. Additionally, Table 6 shows the default hyperparameters were selected on the same MIMIC-III benchmark with CPD as the criterion, which is a selection/leakage concern that could inflate the small TIMING-vs-IG margin (0.366 vs 0.342 at K=50). These concerns weaken the empirical claim, but they are evaluation-validity and data-selection issues rather than circular derivation: the paper also reports independent metrics (AUP/AUR on synthetic data; Acc, CE, Suff, Comp on real data) where TIMING is not uniformly best, and no load-bearing self-citation or imported uniqueness theorem is used. Hence the derivation chain is not circular.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claims depend on the choice of faithfulness metric and the substitution operator, both assumed rather than derived. The method's hyperparameters are hand-selected but robust within the reported range. No new entities are introduced.

free parameters (5)

n (number of segments) = 50 (default)
Chosen by sensitivity analysis (Table 6); default setting (50,10,48) gives best CPD but differences across settings are small.
s_min (minimum segment length) = 10
Minimum segment length; Table 6 shows moderate values work best.
s_max (maximum segment length) = 48
Maximum segment length; larger maxima improve CPD for MIMIC-III.
n_samples (number of alpha steps) = not stated in paper
The algorithm's integration step count is not explicitly reported; affects approximation quality.
K (cumulative masking depth) = 50 and 100
Chosen for reporting; the headline CPD values and rankings depend on K.

assumptions (3)

domain assumption Partial derivatives of the model output along the masked IG paths are bounded and integrable (Prop 4.1 premise)
Invoked in Section 4.3 to justify Fubini interchange; holds for well-behaved classifiers but is not verified for the specific GRU/CNN/Transformer models used.
domain assumption Cumulative prediction difference/preservation measure faithfulness
The claim that CPD/CPP assess attribution quality is asserted in Section 3.2 without external validation against known ground truth on real data.
domain assumption Setting removed points to zero (or the feature mean) is a valid removal operator
Used throughout Section 5; zero substitution may create out-of-distribution inputs, which the paper itself cites as a problem for IG paths.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation." pith.science (2026). https://pith.science/paper/ZSTGOBZW

@misc{pith2026250605035,
  author       = {Pith},
  title        = {Pith review of: TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZSTGOBZW}},
  note         = {Machine review of arXiv:2506.05035}
}

read the original abstract

Recent explainable artificial intelligence (XAI) methods for time series primarily estimate point-wise attribution magnitudes, while overlooking the directional impact on predictions, leading to suboptimal identification of significant points. Our analysis shows that conventional Integrated Gradients (IG) effectively capture critical points with both positive and negative impacts on predictions. However, current evaluation metrics fail to assess this capability, as they inadvertently cancel out opposing feature contributions. To address this limitation, we propose novel evaluation metrics-Cumulative Prediction Difference (CPD) and Cumulative Prediction Preservation (CPP)-to systematically assess whether attribution methods accurately identify significant positive and negative points in time series XAI. Under these metrics, conventional IG outperforms recent counterparts. However, directly applying IG to time series data may lead to suboptimal outcomes, as generated paths ignore temporal relationships and introduce out-of-distribution samples. To overcome these challenges, we introduce TIMING, which enhances IG by incorporating temporal awareness while maintaining its theoretical properties. Extensive experiments on synthetic and real-world time series benchmarks demonstrate that TIMING outperforms existing time series XAI baselines. Our code is available at https://github.com/drumpt/TIMING.

Figures

Figures reproduced from arXiv: 2506.05035 by the authors.

**Figure 1.** An example illustrating how cumulative prediction difference (CPD) improves upon raw prediction difference. While raw difference incorrectly favors a poorly performing method with aligned signs (blue, b) over a perfect method with misaligned signs (red, a), CPD correctly identifies the superior performance of the latter (d vs. c). feature xt,d contributes to increasing model’s prediction score for the chosen clas… view at source ↗

**Figure 2.** Overview of the Temporality-Aware Integrated Gradients (TIMING) framework for improved attribution in time series data. TIMING extends the traditional Integrated Gradients (IG) (Section 4.1) by incorporating temporal dependencies through segment-based random masking to handle disruptions in temporal relationships (Section 4.2). Our framework applies a randomization strategy to compute IG under varied conditions and … view at source ↗

**Figure 3.** Cumulative Prediction Preservation (CPP) comparison of XAI methods on MIMIC-III mortality prediction with zero substitution. Results are averaged over five crossvalidation runs, with 10% random masking of observed points alongside all missing points. Evaluation metrics. As we propose new evaluation metrics for time series XAI—CPD and CPP—and address the limitations of existing metrics, we primarily employ these m… view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Computational efficiency analysis of TIMING and baselines. We report elapsed real time (sec) on a logarithmic scale for all test samples in the MIMIC-III benchmark, alongside CPD with K = 50. Each ordered pair (x, ·) represents per-sample elapsed real-time as the x-c…

**Figure 5.** Figure 5: Cumulative Prediction Preservation (CPP) comparison of XAI methods on MIMIC-III mortality prediction with zero substitution. Results are averaged over five cross-validation runs, with 20% and 40% random masking of observed points alongside all missing points. is finite…

**Figure 6.** Figure 6: Qualitative analysis of input features and attributions extracted from TIMING on the MIMIC-III mortality benchmark (Johnson et al., 2016) for a true positive case where (label = 1, model output = 0.625). Input TIMING [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]

**Figure 7.** Figure 7: Qualitative analysis of input features and attributions extracted from TIMING on the MIMIC-III mortality benchmark (Johnson et al., 2016) for a true positive case where (label = 1, model output = 0.898). Input TIMING [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]

**Figure 8.** Figure 8: Qualitative analysis of input features and attributions extracted from TIMING on the MIMIC-III mortality benchmark (Johnson et al., 2016) for a true negative case where (label = 0, model output = 0.020). Input TIMING [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

**Figure 9.** Figure 9: Qualitative analysis of input features and attributions extracted from TIMING on the MIMIC-III mortality benchmark (Johnson et al., 2016) for a true negative case where (label = 0, model output = 0.081). 18 [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]

**Figure 10.** Figure 10: Qualitative analysis of input features and attributions extracted from TIMING and baselines on the MIMIC-III mortality benchmark (Johnson et al., 2016) for a true positive case where (label = 1, model output = 0.625). 19 [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

TimeSAE: Causal Sparse Decoding for Faithful Explanations of Black-Box Time Series Models
cs.LG 2026-01 conditional novelty 6.0 of 10

TimeSAE trains a sparse autoencoder with counterfactual and consistency losses to explain black-box time series predictions, claiming better faithfulness and out-of-distribution robustness than eight baselines.
Attribution and Uncertainty Behavior of Learned Residual Gyro Correction for Gyro-Stellar Estimation
cs.LG 2026-07 conditional novelty 4.0 of 10

In learned residual gyro correction, ensemble epistemic uncertainty separates nominal from structured sensor perturbations more clearly than heteroscedastic aleatoric uncertainty, while attributions stay gyro-dominate...

Reference graph

Works this paper leans on

47 extracted references · 44 canonical work pages · cited by 2 Pith papers

[1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...
[2]

G., Lehnertz, K., Mormann, F., Rieke, C., David, P., and Elger, C

Andrzejak, R. G., Lehnertz, K., Mormann, F., Rieke, C., David, P., and Elger, C. E. Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state. Physical Review E, 2001

work page 2001
[3]

S., Flunkert, V., Wang, Y., Maddix, D., Turkmen, C., Gasthaus, J., Bohlke-Schneider, M., Salinas, D., Stella, L., et al

Benidis, K., Rangapuram, S. S., Flunkert, V., Wang, Y., Maddix, D., Turkmen, C., Gasthaus, J., Bohlke-Schneider, M., Salinas, D., Stella, L., et al. Deep learning for time series forecasting: Tutorial and literature survey. ACM Computing Surveys, 2022

work page 2022
[4]

Elevated admission lactate levels in the emergency department are associated with increased 30-day mortality in non-trauma critically ill patients

Bernhard, M., D \"o ll, S., Kramer, A., Weidhase, L., Hartwig, T., Petros, S., and Gries, A. Elevated admission lactate levels in the emergency department are associated with increased 30-day mortality in non-trauma critically ill patients. Scandinavian Journal of Trauma, Resuscitation and Emergency Medicine, 2020

work page 2020
[5]

Time series deconfounder: Estimating treatment effects over time in the presence of hidden confounders

Bica, I., Alaa, A., and Van Der Schaar, M. Time series deconfounder: Estimating treatment effects over time in the presence of hidden confounders. In International Conference on Machine Learning, 2020

work page 2020
[6]

and Carlberg, B

Brunstr \"o m, M. and Carlberg, B. Association of blood pressure lowering with mortality and cardiovascular disease across blood pressure levels: A systematic review and meta-analysis. JAMA Internal Medicine, 2018

work page 2018
[7]

Chattopadhay, A., Sarkar, A., Howlader, P., and Balasubramanian, V. N. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2018

work page 2018
[8]

Timemil: Advancing multivariate time series classification via a time-aware multiple instance learning

Chen, X., Qiu, P., Zhu, W., Li, H., Wang, H., Sotiras, A., Wang, Y., and Razi, A. Timemil: Advancing multivariate time series classification via a time-aware multiple instance learning. In International Conference on Machine Learning, 2024

work page 2024

Show all 47 references

[9]

Interpretable machine learning: A guide for making black box models explainable

Christoph, M. Interpretable machine learning: A guide for making black box models explainable. Leanpub, 2020

2020
[10]

Empirical evaluation of gated recurrent neural networks on sequence modeling

Chung, J., Gulcehre, C., Cho, K., and Bengio, Y. Empirical evaluation of gated recurrent neural networks on sequence modeling. NIPS Workshop on Deep Learning and Representation Learning, 2014

2014
[11]

and Van Der Schaar, M

Crabb \'e , J. and Van Der Schaar, M. Explaining time series predictions with dynamic masks. In International Conference on Machine Learning, 2021

2021
[12]

A survey of the state of explainable ai for natural language processing

Danilevsky, M., Qian, K., Aharonov, R., Katsis, Y., Kawas, B., and Sen, P. A survey of the state of explainable ai for natural language processing. In Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th In...

2020
[13]

A., Bagnall, A., Kamgar, K., Yeh, C.-C

Dau, H. A., Bagnall, A., Kamgar, K., Yeh, C.-C. M., Zhu, Y., Gharghabi, S., Ratanamahatana, C. A., and Keogh, E. The ucr time series archive. IEEE/CAA Journal of Automatica Sinica, 2019

2019
[14]

Dick, K., Russell, L., Souley Dosso, Y., Kwamena, F., and Green, J. R. Deep learning for critical infrastructure resilience. Journal of Infrastructure Systems, 2019

2019
[15]

Inherently interpretable time series classification via multiple instance learning

Early, J., Cheung, G., Cutajar, K., Xie, H., Kandola, J., and Twomey, N. Inherently interpretable time series classification via multiple instance learning. In International Conference on Learning Representations, 2024

2024
[16]

Learning perturbations to explain time series predictions

Enguehard, J. Learning perturbations to explain time series predictions. In International Conference on Machine Learning, 2023

2023
[17]

Group, S. R. A randomized trial of intensive versus standard blood-pressure control. New England Journal of Medicine, 2015

2015
[18]

A., Gunady, M., Corrada Bravo, H., and Feizi, S

Ismail, A. A., Gunady, M., Corrada Bravo, H., and Feizi, S. Benchmarking deep learning interpretability in time series predictions. Conference on Neural Information Processing Systems, 2020

2020
[19]

E., Pollard, T

Johnson, A. E., Pollard, T. J., Shen, L., Lehman, L.-w. H., Feng, M., Ghassemi, M., Moody, B., Szolovits, P., Anthony Celi, L., and Mark, R. G. Mimic-iii, a freely accessible critical care database. Scientific data, 2016

2016
[20]

Krizhevsky, A., Sutskever, I., and Hinton, G. E. Imagenet classification with deep convolutional neural networks. Conference on Neural Information Processing Systems, 2012

2012
[21]

K., Rooke, C., Smith, J., Zuberi, S., and Volkovs, M

Leung, K. K., Rooke, C., Smith, J., Zuberi, S., and Volkovs, M. Temporal dependencies in feature importance for time series prediction. In International Conference on Learning Representations, 2023

2023
[22]

Timex++: Learning time-series explanations with information bottleneck

Liu, Z., Wang, T., Shi, J., Zheng, X., Chen, Z., Song, L., Dong, W., Obeysekera, J., Shirani, F., and Luo, D. Timex++: Learning time-series explanations with information bottleneck. In International Conference on Machine Learning, 2024 a

2024
[23]

Explaining time series via contrastive and locally sparse perturbations

Liu, Z., ZHANG, Y., Wang, T., Wang, Z., Luo, D., Du, M., Wu, M., Wang, Y., Chen, C., Fan, L., et al. Explaining time series via contrastive and locally sparse perturbations. In International Conference on Learning Representations, 2024 b

2024
[24]

Lundberg, S. M. and Lee, S.-I. A unified approach to interpreting model predictions. Conference on Neural Information Processing Systems, 30, 2017

2017
[25]

Ma, H., Lin, S., Xie, Y., Mo, S., Huang, Q., Ge, H., Shi, Z., Li, S., and Zhou, D. Association between bun/creatinine ratio and the risk of in-hospital mortality in patients with trauma-related acute respiratory distress syndrome: a single-centre retrospective cohort from the ...

2023
[26]

Deep learning methods in transportation domain: a review

Nguyen, H., Kieu, L.-M., Wen, T., and Cai, C. Deep learning methods in transportation domain: a review. IET Intelligent Transport Systems, 2018

2018
[27]

Encoding time-series explanations through self-supervised model behavior consistency

Queen, O., Hartvigsen, T., Koker, T., He, H., Tsiligkaridis, T., and Zitnik, M. Encoding time-series explanations through self-supervised model behavior consistency. Conference on Neural Information Processing Systems, 2024

2024
[28]

S., Seeger, M

Rangapuram, S. S., Seeger, M. W., Gasthaus, J., Stella, L., Wang, Y., and Januschowski, T. Deep state space models for time series forecasting. Conference on Neural Information Processing Systems, 2018

2018
[29]

S., Werner, L

Rangapuram, S. S., Werner, L. D., Benidis, K., Mercado, P., Gasthaus, J., and Januschowski, T. End-to-end learning of coherent probabilistic forecasts for hierarchical time series. In International Conference on Machine Learning, 2021

2021
[30]

and Stricker, D

Reiss, A. and Stricker, D. Introducing a new benchmarked dataset for activity monitoring. In International Symposium on Wearable Computers, 2012

2012
[31]

T., Singh, S., and Guestrin, C

Ribeiro, M. T., Singh, S., and Guestrin, C. ``why should i trust you?'' explaining the predictions of any classifier. In ACM SIGKDD international conference on knowledge discovery and data mining, 2016

2016
[32]

S., Hughes, M

Ross, A. S., Hughes, M. C., and Doshi-Velez, F. Right for the right reasons: Training differentiable models by constraining their explanations. In International Joint Conference on Artificial Intelligence, 2017

2017
[33]

Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead

Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 2019

2019
[34]

R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., and Batra, D

Selvaraju, R. R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., and Batra, D. Grad-cam: Visual explanations from deep networks via gradient-based localization. In IEEE/CVF International Conference on Computer Vision, 2017

2017
[35]

Simulated boiler data for fault detection and classification, 2019

Shohet, R., Kandil, M., and McArthur, J. Simulated boiler data for fault detection and classification, 2019

2019
[36]

Learning important features through propagating activation differences

Shrikumar, A., Greenside, P., and Kundaje, A. Learning important features through propagating activation differences. In International Conference on Machine Learning, 2017

2017
[37]

Spatial-temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data forecasting

Song, C., Lin, Y., Guo, S., and Wan, H. Spatial-temporal synchronous graph convolutional networks: A new framework for spatial-temporal network data forecasting. In AAAI conference on artificial intelligence, 2020

2020
[38]

Sukkar, R., Katz, E., Zhang, Y., Raunig, D., and Wyman, B. T. Disease progression modeling using hidden markov models. In International Conference of the IEEE Engineering in Medicine and Biology Society, 2012

2012
[39]

Axiomatic attribution for deep networks

Sundararajan, M., Taly, A., and Yan, Q. Axiomatic attribution for deep networks. In International Conference on Machine Learning, 2017

2017
[40]

A., Szolovits, P., and Ghassemi, M

Suresh, H., Hunt, N., Johnson, A., Celi, L. A., Szolovits, P., and Ghassemi, M. Clinical intervention prediction and understanding using deep networks. In Machine Learning for Healthcare, 2017

2017
[41]

Impact of blood urea nitrogen to creatinine ratio on mortality and morbidity in hemodialysis patients: The q-cohort study

Tanaka, S., Ninomiya, T., Taniguchi, M., Tokumoto, M., Masutani, K., Ooboshi, H., Kitazono, T., and Tsuruya, K. Impact of blood urea nitrogen to creatinine ratio on mortality and morbidity in hemodialysis patients: The q-cohort study. Scientific reports, 2017

2017
[42]

K., and Goldenberg, A

Tonekaboni, S., Joshi, S., Campbell, K., Duvenaud, D. K., and Goldenberg, A. What went wrong and when? instance-wise feature importance for time-series black-box models. Conference on Neural Information Processing Systems, 2020

2020
[43]

F., Hadjout, D., Sebaa, A., Mart \' nez- \'A lvarez, F., and Troncoso, A

Torres, J. F., Hadjout, D., Sebaa, A., Mart \' nez- \'A lvarez, F., and Troncoso, A. Deep learning for time series forecasting: a survey. Big Data, 2021

2021
[44]

N., Kaiser, ., and Polosukhin, I

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, ., and Polosukhin, I. Attention is all you need. Conference on Neural Information Processing Systems, 2017

2017
[45]

H., and Lighthall, G

Villar, J., Short, J. H., and Lighthall, G. Lactate predicts both short-and long-term mortality in patients with and without sepsis. Infectious Diseases: Research and Treatment, 2019

2019
[46]

Interpretation of time-series deep models: A survey

Zhao, Z., Shi, Y., Wu, S., Yang, F., Song, W., and Liu, N. Interpretation of time-series deep models: A survey. arXiv preprint arXiv:2305.14582, 2023

2023 arXiv
[47]

Learning deep features for discriminative localization

Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., and Torralba, A. Learning deep features for discriminative localization. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016

2016

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

[2] [2]

G., Lehnertz, K., Mormann, F., Rieke, C., David, P., and Elger, C

Andrzejak, R. G., Lehnertz, K., Mormann, F., Rieke, C., David, P., and Elger, C. E. Indications of nonlinear deterministic and finite-dimensional structures in time series of brain electrical activity: Dependence on recording region and brain state. Physical Review E, 2001

work page 2001

[3] [3]

S., Flunkert, V., Wang, Y., Maddix, D., Turkmen, C., Gasthaus, J., Bohlke-Schneider, M., Salinas, D., Stella, L., et al

Benidis, K., Rangapuram, S. S., Flunkert, V., Wang, Y., Maddix, D., Turkmen, C., Gasthaus, J., Bohlke-Schneider, M., Salinas, D., Stella, L., et al. Deep learning for time series forecasting: Tutorial and literature survey. ACM Computing Surveys, 2022

work page 2022

[4] [4]

Elevated admission lactate levels in the emergency department are associated with increased 30-day mortality in non-trauma critically ill patients

Bernhard, M., D \"o ll, S., Kramer, A., Weidhase, L., Hartwig, T., Petros, S., and Gries, A. Elevated admission lactate levels in the emergency department are associated with increased 30-day mortality in non-trauma critically ill patients. Scandinavian Journal of Trauma, Resuscitation and Emergency Medicine, 2020

work page 2020

[5] [5]

Time series deconfounder: Estimating treatment effects over time in the presence of hidden confounders

Bica, I., Alaa, A., and Van Der Schaar, M. Time series deconfounder: Estimating treatment effects over time in the presence of hidden confounders. In International Conference on Machine Learning, 2020

work page 2020

[6] [6]

and Carlberg, B

Brunstr \"o m, M. and Carlberg, B. Association of blood pressure lowering with mortality and cardiovascular disease across blood pressure levels: A systematic review and meta-analysis. JAMA Internal Medicine, 2018

work page 2018

[7] [7]

Chattopadhay, A., Sarkar, A., Howlader, P., and Balasubramanian, V. N. Grad-cam++: Generalized gradient-based visual explanations for deep convolutional networks. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2018

work page 2018

[8] [8]

Timemil: Advancing multivariate time series classification via a time-aware multiple instance learning

Chen, X., Qiu, P., Zhu, W., Li, H., Wang, H., Sotiras, A., Wang, Y., and Razi, A. Timemil: Advancing multivariate time series classification via a time-aware multiple instance learning. In International Conference on Machine Learning, 2024

work page 2024