Pith. sign in

REVIEW 4 major objections 5 minor 80 references

TimeSAE uses a sparse autoencoder with counterfactual concept interventions to produce post-hoc explanations of black-box time series models that remain faithful and robust under distribution shift, backed by a theoretical guarantee that ap

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 10:32 UTC pith:KGMXZPFY

load-bearing objection TimeSAE is a genuinely new SAE-based post-hoc explainer for time series with solid synthetic results, but its headline real-data superiority rests on under-specified ground-truth masks and a theorem that assumes its own conclusion. the 4 major comments →

arxiv 2601.09776 v2 pith:KGMXZPFY submitted 2026-01-14 cs.LG

TimeSAE: Causal Sparse Decoding for Faithful Explanations of Black-Box Time Series Models

Khalid Oublal , Quentin Bouniot , Qi Gan , Stephan Cl\'emen\c{c}on , Zeynep Akata This is my paper
classification cs.LG
keywords time series explainabilitysparse autoencodercounterfactual explanationout-of-distribution generalizationconcept learningJumpReLUfaithfulnesscausal effect
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper introduces TimeSAE, a post-hoc explainer for black-box time series models. It learns a sparse set of interpretable concepts via a JumpReLU sparse autoencoder, then generates explanations by intervening on those concepts to produce counterfactuals. The central claim is that these concept-based counterfactuals are more faithful and more stable under distribution shift than mask-based attribution methods such as TimeX and TimeX++. A supporting theorem states that, when reconstruction and counterfactual approximation errors are small, the method provably preserves the ordering of causal effects among concepts. If correct, TimeSAE offers a model-agnostic way to explain time series classifiers and forecasters without needing access to the model's internals.

Core claim

The core discovery is that a sparse autoencoder, trained with a contrastive counterfactual loss and a compositional consistency loss, can serve as a faithful explainer for arbitrary time series models. TimeSAE encodes an input into sparse non-negative concept activations via JumpReLU, decodes them back into 'explanation-embedded' instances that preserve the black-box model's label, and intervenes on individual concepts to create approximate counterfactuals. The authors prove (Theorem 1) that if the combined reconstruction and counterfactual approximation error is below half the gap between true causal effects, then the ordering of causal effects is preserved, meaning concepts ranked as more

What carries the argument

The load-bearing component is the sparse autoencoder with JumpReLU activation: an overcomplete dictionary of concept directions with per-feature learnable thresholds that avoids 'dead' concepts. Interventions on these concepts generate counterfactual instances via the decoder, and the decoder's functional-ANOVA decomposition exposes first- and higher-order concept interactions as saliency masks. Training couples three objectives: reconstruction fidelity and sparsity, a label-fidelity term that keeps reconstructed inputs close in the black-box model's output, an InfoNCE counterfactual loss that enforces ordering of causal effects, and a compositional consistency loss that makes the encoder in

Load-bearing premise

The load-bearing premise is that the ground-truth saliency maps used to score the methods are correct: for the ECG data the scoring rule is stated (the QRS interval), but for the PAM dataset and for the regression tasks no explicit protocol defines which features 'cause prediction label changes'—and regression outputs have no label changes—so all AUPRC rankings depend on masks whose validity is asserted rather than demonstrated; the theoretical guarantee also assumes a small

What would settle it

A concrete test would be to construct a synthetic time series with known generative factors (e.g., a 'primary shape' and a 'secondary frequency' as in FreqShapes), compute true causal effects by intervening on those factors, then compare them to TimeSAE's approximate effects across many intervention pairs; if a substantial fraction of pairs violate the ordering when the measured approximation error is kept below half the causal gap, Theorem 1's guarantee is empirically falsified. On the evaluation side, replacing the undefined ground-truth masks for PAM and regression datasets with an explicit

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Post-hoc explanations of time series black-box models can be generated without any access to model internals, requiring only input-output queries.
  • Explanations remain usable when the explainer is applied to data from a shifted distribution, as demonstrated on an ETTh1-to-ETTh2 transfer setup.
  • The framework extends to forecasting and regression tasks, not just classification, including pretrained large models like TimeGPT and Chronos.
  • Concept-based explanations can be interpreted as ordered causal effects, so practitioners can rank which signal components to act on.
  • The new EliteLJ dataset, with human-pose kinematics from long jumps and expert annotations, provides a benchmark for future time-series explanation research.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • As an editorial inference, the same sparse-dictionary machinery could be ported to other black-box explainability settings (e.g., audio or multivariate sensors) because the framework only assumes a queryable model and an encoder-decoder, not model internals.
  • As an editorial inference, the guarantee in Theorem 1 is conditional on an unmeasurable error term; a practical workaround is to report, on synthetic benchmarks, the distribution of approximation errors alongside the fraction of intervention pairs whose ordering is preserved, which the paper does only partially.
  • As an editorial inference, the functional-ANOVA decoder could double as a diagnostic tool for identifying non-additive or spurious interactions in black-box models, an application the paper does not pursue.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes TimeSAE, a post-hoc, model-agnostic explainer for black-box time series classifiers and regressors. TimeSAE trains a sparse autoencoder with a JumpReLU activation to decompose inputs into learned concepts, and adds: (i) a counterfactual InfoNCE loss intended to enforce faithfulness, (ii) a compositional-consistency loss to encourage encoder/decoder invertibility out-of-distribution, and (iii) a label-fidelity loss. The decoder is given a functional-ANOVA-like interaction structure with Bernoulli-style masks to expose concept interactions. The authors claim a theoretical order-faithfulness guarantee for approximate counterfactuals (Theorem 1), and support the method with experiments on two synthetic and several real-world datasets against eight or more baselines, reporting AUPRC/AUP/AUR, faithfulness Fx, distributional alignment, and an OOD study. A new dataset, EliteLJ, is introduced.

Significance. If the empirical claims were fully supported, TimeSAE would be a useful contribution: it is model-agnostic, post-hoc, operates in concept space rather than by input masking, and the code and dataset are publicly released. The paper also attempts to connect sparse-autoencoder counterfactuals to order-faithfulness, which is a worthwhile theoretical target. The evaluation is extensive in breadth (multiple black-box predictors, including pretrained models, multiple datasets, ablations, and an OOD transfer study), and the ablation isolating the counterfactual and consistency losses is informative. However, the central empirical superiority claim currently rests on ground-truth masks whose construction is unspecified for several real-world datasets, and the faithfulness metric Fx is closely aligned with TimeSAE's own training objective. The theoretical guarantee is conditional on an unverified small-error assumption. These issues are load-bearing: until they are addressed, the paper's headline conclusions are not independently testable.

major comments (4)
  1. [§4, 'Explanation Evaluation'; Appendix B.1] The AUPRC/AUP/AUR rankings in Figures 2–3 and Table 9 require binary ground-truth saliency masks for every dataset. The protocol is explicit only for the synthetic datasets and ECG (spike positions; QRS interval). For PAM no mask-construction protocol is given anywhere in the main text or appendix. For the regression datasets ETTh-1, ETTh-2, and EliteLJ, the paper says salient features are 'features causing prediction label changes,' but regression outputs are continuous, so this definition is vacuous without a discretization or thresholding rule. Appendix B.1 describes dataset statistics but not how masks were obtained. Because every real-world AUPRC comparison depends on these masks, the claimed superiority over TimeX++ and CounTS on real-world data is not yet established. Please specify the mask protocol for each dataset, or replace the real-world ranking evaluation with a protocol th
  2. [§4, 'Faithfulness Evaluation'; Eq. (10)] The faithfulness metric Fx = ||f(x) − f(ẽx−)||² is computed by removing the components a method flags and measuring the change in f's output. TimeSAE's training loss includes L_cf and L_label-fidelity, both of which directly reward exactly this kind of output change: Eq. (8) pulls together counterfactual reconstructions that alter f's output, and L_label-fidelity penalizes changes between f(x) and f(g(E(x))). Thus Fx is not an independent evaluation for TimeSAE; it is essentially the training objective. The comparison in Table 1 therefore conflates optimization with faithfulness. To support the faithfulness claim, the authors should use a separate faithfulness measure (e.g., a removal test over ground-truth concepts, or a metric not used in training) or demonstrate that TimeSAE's Fx advantage survives when baselines are also optimized with the same objective.
  3. [Theorem 1, Eq. (5); Appendix A.2] The theorem's conclusion is conditional on the assumption E|f(ẽx_cf) − y_cf| ≤ ε_cf with ε_cf small, where y_cf is the 'true' counterfactual label. The paper does not show that TimeSAE achieves this bound, and in general y_cf requires a causal model of the black box that is unavailable. In the proof, 'there exists a sufficiently small ε_cf' is tautological once Eq. (5) is assumed, because the true causal gap δ is fixed and one can always choose ε_cf < δ/2; the theorem does not establish that the SAE's approximation error is actually that small. The statement also relies on the informal assumption f(ẽx) ≈ f(x) without a quantitative reconstruction-error bound in the theorem. Please reframe the result as a consistency/robustness statement with explicit conditions that can be checked empirically, or provide a bound on ε_cf in terms of quantities the method actually controls.
  4. [§4.1; Abstract; §5] The abstract and conclusion claim that TimeSAE provides 'more faithful and robust explanations' than baselines, but §4.1 reports that TimeSAE is statistically best on 'several datasets' and that 'TimeX++ and CounTS remain competitive.' The significance-testing summary is weaker than the headline claim. If the full Table 9 shows TimeSAE ranked first on every row, then the t-test narrative undercuts that table; if the t-test is the correct summary, then the abstract should be qualified. Please report confidence intervals or effect sizes for the main comparisons and align the abstract's claims with the statistical evidence.
minor comments (5)
  1. [Table 4; §B.1.1] SeqComb-UV is described as univariate in the text but listed with D=10 and task type 'Multivariate' in Table 4. Please clarify the dimension and naming.
  2. [§3.4, Eq. (14)] The Bernoulli masks m_k^(j) are introduced notationally, but the paper never states how p0 is set, whether the masks are learned or fixed, or how the functional-ANOVA decomposition is used in the experiments. This is a substantial part of the proposed decoder, yet it is not evaluated or ablated.
  3. [Eq. (8)] The InfoNCE loss is described as making 'positive pairs' counterfactuals from the same intervention and 'negative pairs' from different interventions. It is unclear how this enforces CaCE ordering without access to ground-truth causal effects. Please clarify the pairing rule.
  4. [§B.3.2] Chronos is described as having '188 billion parameters.' The published Chronos models are far smaller; please verify the stated parameter count and specify which variant was used.
  5. [Figure 2 caption] The caption says 'all datasets' but the figure panels show only FreqShapes and SeqComb-UV. The legend ordering ('worst to best') is also hard to read; consider a clearer layout.

Circularity Check

0 steps flagged

No load-bearing circularity: the central derivation is conditional and externally benchmarked; score reflects only minor non-load-bearing self-citations.

full rationale

The closest candidate for circularity is the faithfulness evaluation: Fx removes concept components and measures ||f(x)-f(g(c-))||^2, while Eq. (10) includes a label-fidelity term and Eq. (8) a counterfactual InfoNCE loss. However, the counterfactual loss is a contrastive objective over intervention pairs, not an objective that directly maximizes this removal effect, and Fx is applied identically to all baselines; there is no equation-level identity between L_cf and Fx, so the concern remains a fairness/validity issue rather than a circular reduction. Theorem 1 is explicitly conditional: Eq. (5) assumes the approximate counterfactual label error is small, and the proof (Appendix A.2) derives the ordering statement by a triangle-inequality argument. The assumption is not the same as the conclusion (order preservation), and the paper validates the assumption on FreqShapes by manipulating ground-truth generative factors (Figure 7, rho=0.94), providing external support. The real-data AUPRC evaluation relies on unspecified ground-truth masks for PAM and regression datasets, but that is an empirical validity threat, not circularity. The few self-citations (Pach et al. 2025; Parekh et al. 2025; Gan et al. 2024) support background sparsity/monosemanticity claims, related-work categorization, and annotation tooling, and are not load-bearing for the central derivation. Hence no significant circularity, only minor non-load-bearing self-citation.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 3 invented entities

The central empirical claim rests on the validity of ground-truth masks, a fairness-neutral faithfulness protocol, and per-dataset hyperparameter selection; the theoretical claim rests on an unverifiable bound on counterfactual approximation error (Eq. 5) plus transfer of identifiability results from object-centric learning. The method introduces no physical entities, but it asserts causal manipulability of learned concepts without a causal graph. Many entries are per-dataset tuning choices rather than physics-style constants, but they are load-bearing because the reported wins are small.

free parameters (7)
  • r (dictionary size ratio) = 1.4–1.7 per dataset (Tables 7–8)
    Chosen per dataset by test-set ablation (Fig. 10); the claimed 'optimal r ≈ 1.5–1.7' rule is itself derived from the reported tuning curves.
  • α (consistency weight) = 0.8–1.0 per dataset (Tables 7–8)
    Tuned per dataset; Fig. 14 shows sensitivity of AUPRC/AUP/AUR to α.
  • λ (counterfactual weight) = 0.8–1.0 per dataset (Tables 7–8)
    Tuned per dataset; drives the faithfulness effect that is the paper's headline claim.
  • η (sparsity coefficient) = not tabulated per dataset
    Sparsity–reconstruction tradeoff tuned in Fig. 5-c; JumpReLU thresholds φ are learned per feature.
  • τ (InfoNCE temperature) = not reported
    Controls sharpness of the contrastive similarity distribution in Eq. (8); not listed in any hyperparameter table.
  • γ_max (TopK scheduler) = 1–12 per dataset (Table 7)
    Initial TopK sparsity level for TimeSAE-TopK; scheduled down to 1 per Eq. (19).
  • LR, dropout, batch size, weight decay, epochs = per-dataset (Tables 7–8)
    Standard training knobs, all tuned per dataset with no nested validation protocol described.
axioms (6)
  • domain assumption Wiedemer et al. (2024) slot-identifiability conditions transfer from object-centric learning to time series concept spaces.
    Invoked in §3.3 to justify that the compositional consistency loss Lcc yields OOD-consistent encoder inversion; the cited theorem's assumptions are not restated and may fail for non-factored time series data.
  • domain assumption Latent concepts are causally manipulable: do(c_k = c'_k) on SAE latents is a well-defined intervention under CaCE.
    §3.2 Definition 2 treats learned dictionary features as if they support Pearl-style interventions; no causal graph or identifiability argument for the latents is provided.
  • ad hoc to paper Eq. (5): E|f(ẽ_cf) − y_cf| ≤ ε_cf with ε_cf small.
    Theorem 1's entire force is conditional on this unverifiable bound; y_cf is the 'true counterfactual label' of a black box, which presumes knowledge of its exact causal input-output structure.
  • domain assumption f(ẽ) ≈ f(x) for all x in D (label fidelity).
    Theorem 1 premises this; in practice it is encouraged by training with L_label-fidelity rather than proven to hold.
  • ad hoc to paper The decoder is well represented by a functional-ANOVA decomposition with independent Bernoulli masks m_k^(j).
    §3.4 and Eq. (14) claim this decomposition, but the implementation (Table 6) is FC + attention + upsampling + conv stacks with no described ANOVA structure; the masks are never validated in experiments.
  • domain assumption ETTh1→ETTh2 is a valid OOD benchmark for explanation generalization.
    Table 2-a treats the two ETT datasets as an OOD pair; they differ by country/seasonal patterns, but no distributional or causal characterization of the shift is given.
invented entities (3)
  • Interpretable concept dictionary (rows of M with JumpReLU thresholds) independent evidence
    purpose: The explanation units; interventions on them define the counterfactuals used to measure faithfulness.
    Concepts are testable against external ground-truth saliency masks (synthetic data, ECG QRS) via AUPRC, giving a falsifiable handle.
  • Bernoulli interaction masks m_k^(j) over ANOVA ψ_k terms no independent evidence
    purpose: Attribute decoder output to concept orders and interactions.
    No experiment isolates, validates, or visualizes these masks; they are introduced in §3.4 and never connected to results.
  • EliteLJ dataset (elite long-jump pose sequences with phase labels) independent evidence
    purpose: New benchmark for time series explanation evaluation with expert annotations.
    Claims public release and manual pose corrections; however, the ground-truth explanation masks for AUPRC evaluation are not specified.

pith-pipeline@v1.3.0-alltime-deepseek · 34136 in / 24072 out tokens · 226963 ms · 2026-08-03T10:32:27.156369+00:00 · methodology

0 comments
read the original abstract

As black box models and pretrained models gain traction in time series applications, understanding and explaining their predictions becomes increasingly vital, especially in high-stakes domains where interpretability and trust are essential. However, most of the existing methods involve only in-distribution explanation, and do not generalize outside the training support, which requires the learning capability of generalization. In this work, we aim to provide a framework to explain black-box models for time series data through the dual lenses of Sparse Autoencoders (SAEs) and causality. We show that many current explanation methods are sensitive to distributional shifts, limiting their effectiveness in real-world scenarios. Building on the concept of Sparse Autoencoder, we introduce TimeSAE, a framework for black-box model explanation. We conduct extensive evaluations of TimeSAE on both synthetic and real-world time series datasets, comparing it to leading baselines. The results, supported by both quantitative metrics and qualitative insights, show that TimeSAE provides more faithful and robust explanations. Our code is available in an easy-to-use library TimeSAE-Lib: https://oublalkhalid.github.io/TimeSAE/

Figures

Figures reproduced from arXiv: 2601.09776 by Khalid Oublal, Qi Gan, Quentin Bouniot, Stephan Cl\'emen\c{c}on, Zeynep Akata.

Figure 1
Figure 1. Figure 1: Overview of Time Series Sparse Autoencoder ( TimeSAE ): (A) The framework assumes access to a black-box model f and aims to explain its predictions on data x ∈ X by learning an explainer E and a decoder g that decompose the time series into interpretable components. (B) For faithfulness, the sparse autoencoder (E, g) incorporates properties to leverage counterfactual explanations. A contrastive learning lo… view at source ↗
Figure 2
Figure 2. Figure 2: Explanation performance on all datasets and metrics (AUPRC, AUP, AUR). Higher is better. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: AUPRC explanation performance (higher is better) across methods for each dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Spearman correlation (ρ ≈ −0.981) between the Faithfulness metric (Fx) and the Counterfactual Approximation Error (ϵcf ). Theory Validation. To empirically validate our the￾ory, we leverage the existing faithfulness (Fx) metric used in our main results. Furthermore, we inves￾tigate the counterfactual approximation error (ϵcf ) mentioned in Theorem 1. Specifically, we compute the Spearman correlation betwee… view at source ↗
Figure 5
Figure 5. Figure 5: (a) Examples of explanation compared to the ground truth on the FreqShapes dataset. (b) Effects of excluding the Concepts Consistency and Counterfactual term on the Faithfulness metric Fx, with standard deviations shown over 10 runs. Using counterfactuals produces more faithful explanations. (c) Effect of sparsity on reconstruction fidelity. Increasing sparsity generally improves interpretability; however,… view at source ↗
Figure 6
Figure 6. Figure 6: Sparsity efficiency and activation frequency. Left: TimeGPT on EliteLJ. Middle: Chronos on EliteLJ. In both, TimeSAE with JumpReLU outperforms TopK at all L0, showing better sparsity￾fidelity trade-offs. Right: Log-scale activation histogram: TopK spikes near 0% (dead concept), JumpReLU is more distributed; 10 seeds shown in gradient colors. Effectiveness of Counterfactual for Faithful Explanations We eval… view at source ↗
Figure 7
Figure 7. Figure 7: Empirical Analysis of Theorem 1. (a) Scatter plot showing strong correlation between the true causal effects and SAE-estimated effects, confirming order-faithfulness. (b) For a specific pair of interventions, the measured approximation error is smaller than half the true gap, preventing the reversal of causal ordering. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: 24-hour forecast based on a 48-hour history from the ETTh1 dataset. The heatmap [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: 24-hour forecast based on a 48-hour history from the ETTh1 dataset. The heatmap visualizes [PITH_FULL_IMAGE:figures/full_fig_p026_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Impact of the hyperparameter r on model performance. Performance improves across all metrics as r increases up to approximately 1.6 for Transformer, 1.5 for PatchTS and TimeGPT, and 1.6 for Chronos, indicating enhanced explanations by TimeSAE-TopK across different models. Beyond values around 1.9, performance deteriorates, likely due to high sparsity. We note that higher metric values correspond to better… view at source ↗
Figure 11
Figure 11. Figure 11: Visualization Explanation for the Transformer model’s predictions on the FreqShapes [PITH_FULL_IMAGE:figures/full_fig_p030_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Left: Concept activations over training steps using hard Top-K with a fixed high γ value, evaluated on the same fixed validation time series sequence across training steps on the SeqComb-UV dataset, to explain the vanilla Transformer. As training progresses, many concepts exhibit near-zero activations, indicating the emergence of “dead” concepts that stop learning effectively. Right: Concept activations u… view at source ↗
Figure 13
Figure 13. Figure 13: Intuition behind the effect of sparsity. [PITH_FULL_IMAGE:figures/full_fig_p033_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Extended ablation study on the effect of the concept consistency weight [PITH_FULL_IMAGE:figures/full_fig_p033_14.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

80 extracted references · 20 linked inside Pith

  1. [1]

    Can CO2 emissions and energy consumption determine the economic performance of south korea? A time series analysis

    Tomiwa Sunday Adebayo, Abraham Ayobamiji Awosusi, Dervis Kirikkaleli, Gbenga Daniel Akinsola, and Madhy Nyota Mwamba. Can CO2 emissions and energy consumption determine the economic performance of south korea? A time series analysis. Environmental Science and Pollution Research, pp.\ 38969--38984, 2021

    2021

  2. [2]

    Towards robust interpretability with self-explaining neural networks

    David Alvarez Melis and Tommi Jaakkola. Towards robust interpretability with self-explaining neural networks. Advances in neural information processing systems, 31, 2018

    2018

  3. [3]

    Chronos: Learning the language of time series

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebastian Pineda Arango, Shubham Kapoor, et al. Chronos: Learning the language of time series. arXiv preprint arXiv:2403.07815, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    How to explain individual classification decisions

    David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja Hansen, and Klaus-Robert M \"u ller. How to explain individual classification decisions. The Journal of Machine Learning Research, 11: 0 1803--1831, 2010

    2010

  5. [5]

    Discord-based counterfactual explanations for time series classification

    Omar Bahri, Peiyu Li, Soukaina Filali Boubrahimi, and Shah Muhammad Hamdi. Discord-based counterfactual explanations for time series classification. Data Mining and Knowledge Discovery, 38 0 (6): 0 3347--3371, 2024

    2024

  6. [6]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling

    Shaojie Bai, J Zico Kolter, and Vladlen Koltun. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv preprint arXiv:1803.01271, 2018

    Pith/arXiv arXiv 2018

  7. [7]

    Timeshap: Explaining recurrent models through sequence perturbations

    Jo \ a o Bento, Pedro Saleiro, Andr \'e F Cruz, M \'a rio AT Figueiredo, and Pedro Bizarro. Timeshap: Explaining recurrent models through sequence perturbations. In SIGKDD, pp.\ 2565--2573, 2021

    2021

  8. [8]

    o lkopf, Julius Von K\

    Jack Brady, Roland S. Zimmermann, Yash Sharma, Bernhard Sch\" o lkopf, Julius Von K\" u gelgen, and Wieland Brendel. Provably learning object-centric representations. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp.\ 3038--3062. PMLR, 23--29 Jul 2023

    2023

  9. [9]

    Co RTX : Contrastive framework for real-time explanation

    Yu-Neng Chuang, Guanchu Wang, Fan Yang, Quan Zhou, Pushkar Tripathi, Xuanting Cai, and Xia Hu. Co RTX : Contrastive framework for real-time explanation. In ICLR, pp.\ 1--23, 2023

    2023

  10. [10]

    Explaining time series predictions with dynamic masks

    Jonathan Crabb \'e and Mihaela Van Der Schaar. Explaining time series predictions with dynamic masks. In ICML, pp.\ 2166--2177, 2021

    2021

  11. [11]

    Concept activation regions: A generalized framework for concept-based explanations

    Jonathan Crabb \'e and Mihaela van der Schaar. Concept activation regions: A generalized framework for concept-based explanations. Advances in Neural Information Processing Systems, 35: 0 2590--2607, 2022

    2022

  12. [12]

    Sparse autoencoders find highly interpretable features in language models

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Comparative study of machine learning methods for covid-19 transmission forecasting

    Abdelkader Dairi, Fouzi Harrou, Abdelhafid Zeroual, Mohamad Mazen Hittawe, and Ying Sun. Comparative study of machine learning methods for covid-19 transmission forecasting. Journal of biomedical informatics, 118: 0 103791, 2021

    2021

  14. [14]

    A decoder-only foundation model for time-series forecasting

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. A decoder-only foundation model for time-series forecasting. In Forty-first International Conference on Machine Learning, 2024

    2024

  15. [15]

    The ucr time series archive

    Hoang Anh Dau, Anthony Bagnall, Kaveh Kamgar, Chin-Chia Michael Yeh, Yan Zhu, Shaghayegh Gharghabi, Chotirat Ann Ratanamahatana, and Eamonn Keogh. The ucr time series archive. IEEE/CAA Journal of Automatica Sinica, 6 0 (6): 0 1293--1305, 2019

    2019

  16. [16]

    Managing electric flexibility from distributed energy resources: A review of incentives for market design

    Cherrelle Eid, Paul Codani, Yannick Perez, Javier Reneses, and Rudi Hakvoort. Managing electric flexibility from distributed energy resources: A review of incentives for market design. Renewable and Sustainable Energy Reviews, 64: 0 237--247, 2016

    2016

  17. [17]

    Learning perturbations to explain time series predictions

    Joseph Enguehard. Learning perturbations to explain time series predictions. In ICML, pp.\ 9329--9342, 2023

    2023

  18. [18]

    Human pose estimation based biomechanical feature extraction for long jumps

    Qi Gan, Moun \^ m A El-Yacoubi, Eric Fenaux, St \'e phan Cl \'e men c on, et al. Human pose estimation based biomechanical feature extraction for long jumps. In 2024 16th International Conference on Human System Interaction (HSI), pp.\ 1--6. IEEE, 2024

    2024

  19. [19]

    Scaling and evaluating sparse autoencoders

    Leo Gao, Tom Dupr \'e la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. Scaling and evaluating sparse autoencoders. arXiv preprint arXiv:2406.04093, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Timegpt-1

    Azul Garza, Cristian Challu, and Max Mergenthaler-Canseco. Timegpt-1. arXiv preprint arXiv:2310.03589, 2023

    Pith/arXiv arXiv 2023

  21. [21]

    Faithful explanations of black-box nlp models using llm-generated counterfactuals

    Yair Gat, Nitay Calderon, Amir Feder, Alexander Chapanin, Amit Sharma, and Roi Reichart. Faithful explanations of black-box nlp models using llm-generated counterfactuals. arXiv preprint arXiv:2310.00603, 2023

    Pith/arXiv arXiv 2023

  22. [22]

    Patchscope: A unifying framework for inspecting hidden representations of language models

    Asma Ghandeharioun, Avi Caciularu, Adam Pearce, Lucas Dixon, and Mor Geva. Patchscope: A unifying framework for inspecting hidden representations of language models. arXiv preprint arXiv:2401.06102, 2024

    Pith/arXiv arXiv 2024

  23. [23]

    Moment: A family of open time-series foundation models

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. Moment: A family of open time-series foundation models. arXiv preprint arXiv:2402.03885, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Explaining classifiers with causal concept effect (cace)

    Yash Goyal, Amir Feder, Uri Shalit, and Been Kim. Explaining classifiers with causal concept effect (cace). CoRR, abs/1907.07165, 2019. URL http://arxiv.org/abs/1907.07165

    Pith/arXiv arXiv 1907

  25. [25]

    A kernel two-sample test

    Arthur Gretton, Karsten M Borgwardt, Malte J Rasch, Bernhard Sch \"o lkopf, and Alexander Smola. A kernel two-sample test. The journal of machine learning research, 13 0 (1): 0 723--773, 2012

    2012

  26. [26]

    Squeeze-and-excitation networks

    Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 7132--7141, 2018

    2018

  27. [27]

    Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence, 36 0 (7): 0 1325--1339, 2013

    2013

  28. [28]

    Improving deep learning interpretability by saliency guided training

    Aya Abdelsalam Ismail, Hector Corrada Bravo, and Soheil Feizi. Improving deep learning interpretability by saliency guided training. In NeurIPS, pp.\ 26726--26739, 2021

    2021

  29. [29]

    Attention is not explanation

    Sarthak Jain and Byron C Wallace. Attention is not explanation. arXiv preprint arXiv:1902.10186, 2019

    Pith/arXiv arXiv 1902

  30. [30]

    Timing: Temporality-aware integrated gradients for time series explanation

    Hyeongwon Jang, Changhun Kim, and Eunho Yang. Timing: Temporality-aware integrated gradients for time series explanation. arXiv preprint arXiv:2506.05035, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    Multivariate lstm-fcns for time series classification

    Fazle Karim, Somshubra Majumdar, Houshang Darabi, and Samuel Harford. Multivariate lstm-fcns for time series classification. Neural networks, 116: 0 237--245, 2019

    2019

  32. [32]

    AI in healthcare: time-series forecasting using statistical, neural, and ensemble architectures

    Shruti Kaushik, Abhinav Choudhury, Pankaj Kumar Sheron, Nataraj Dasgupta, Sayee Natarajan, Larry A Pickett, and Varun Dutt. AI in healthcare: time-series forecasting using statistical, neural, and ensemble architectures. Frontiers in Big Data, 3: 0 4, 2020

    2020

  33. [33]

    Concept bottleneck models

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pp.\ 5338--5348. PMLR, 2020

    2020

  34. [34]

    Temporal dependencies in feature importance for time series prediction

    Kin Kwan Leung, Clayton Rooke, Jonathan Smith, Saba Zuberi, and Maksims Volkovs. Temporal dependencies in feature importance for time series prediction. In ICLR, pp.\ 1--18, 2023

    2023

  35. [35]

    Cels: Counterfactual explanations for time series data via learned saliency maps

    Peiyu Li, Omar Bahri, Souka \" na Filali Boubrahimi, and Shah Muhammad Hamdi. Cels: Counterfactual explanations for time series data via learned saliency maps. In 2023 IEEE International Conference on Big Data (BigData), pp.\ 718--727. IEEE, 2023

    2023

  36. [36]

    Reliable time series counterfactual explanations guided by shapedba

    Peiyu Li, Pouya Hosseinzadeh, Omar Bahri, Souka \" na Filali Boubrahimi, and Shah Muhammad Hamdi. Reliable time series counterfactual explanations guided by shapedba. In 2024 IEEE International Conference on Big Data (BigData), pp.\ 1574--1579. IEEE, 2024

    2024

  37. [37]

    Time-series forecasting for out-of-distribution generalization using invariant learning

    Haoxin Liu, Harshavardhan Kamarthi, Lingkai Kong, Zhiyuan Zhao, Chao Zhang, and B Aditya Prakash. Time-series forecasting for out-of-distribution generalization using invariant learning. In Proceedings of the 41st International Conference on Machine Learning, pp.\ 31312--31325, 2024 a

    2024

  38. [38]

    itransformer: Inverted transformers are effective for time series forecasting

    Yong Liu, Tengge Hu, Haoran Zhang, Haixu Wu, Shiyu Wang, Lintao Ma, and Mingsheng Long. itransformer: Inverted transformers are effective for time series forecasting. arXiv preprint arXiv:2310.06625, 2023

    Pith/arXiv arXiv 2023

  39. [39]

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

    Zichuan Liu, Tianchun Wang, Jimeng Shi, Xu Zheng, Zhuomin Chen, Lei Song, Wenqian Dong, Jayantha Obeysekera, Farhad Shirani, and Dongsheng Luo. Timex++: Learning time-series explanations with information bottleneck. arXiv preprint arXiv:2405.09308, 2024 b

    Pith/arXiv arXiv 2024

  40. [40]

    Explaining time series via contrastive and locally sparse perturbations

    Zichuan Liu, Yingying Zhang, Tianchun Wang, Zefan Wang, Dongsheng Luo, Mengnan Du, Min Wu, Yi Wang, Chunlin Chen, Lunting Fan, and Qingsong Wen. Explaining time series via contrastive and locally sparse perturbations. In ICLR, pp.\ 1--21, 2024 c

    2024

  41. [41]

    A unified approach to interpreting model predictions

    Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems, pp.\ 4765--4774, 2017

    2017

  42. [42]

    K-sparse autoencoders

    Alireza Makhzani and Brendan Frey. K-sparse autoencoders. arXiv preprint arXiv:1312.5663, 2013

    Pith/arXiv arXiv 2013

  43. [43]

    Neural decomposition: Functional anova with variational autoencoders

    Kaspar M \"a rtens and Christopher Yau. Neural decomposition: Functional anova with variational autoencoders. In International Conference on Artificial Intelligence and Statistics, pp.\ 2917--2927. PMLR, 2020

    2020

  44. [44]

    Interpretable and generalizable graph learning via stochastic attention mechanism

    Siqi Miao, Mia Liu, and Pan Li. Interpretable and generalizable graph learning via stochastic attention mechanism. In ICML, pp.\ 15524--15543, 2022

    2022

  45. [45]

    Distributed representations of words and phrases and their compositionality

    Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013

    2013

  46. [46]

    The impact of the mit-bih arrhythmia database

    George B Moody and Roger G Mark. The impact of the mit-bih arrhythmia database. IEEE engineering in medicine and biology magazine, 20 0 (3): 0 45--50, 2001

    2001

  47. [47]

    Sparse autoencoder

    Andrew Ng. Sparse autoencoder. http://web.stanford.edu/class/cs294a/sparseAutoencoder.pdf, 2011. CS294A Lecture notes

    2011

  48. [48]

    A time series is worth 64 words: Long-term forecasting with transformers

    Yuqi Nie, Nam H Nguyen, Phanwadee Sinthong, and Jayant Kalagnanam. A time series is worth 64 words: Long-term forecasting with transformers. arXiv preprint arXiv:2211.14730, 2022

    Pith/arXiv arXiv 2022

  49. [49]

    Label-free concept bottleneck models

    Tuomas Oikarinen, Subhro Das, Lam M Nguyen, and Tsui-Wei Weng. Label-free concept bottleneck models. arXiv preprint arXiv:2304.06129, 2023

    Pith/arXiv arXiv 2023

  50. [50]

    Representation learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  51. [51]

    Sparse autoencoders learn monosemantic features in vision-language models

    Mateusz Pach, Shyamgopal Karthik, Quentin Bouniot, Serge Belongie, and Zeynep Akata. Sparse autoencoders learn monosemantic features in vision-language models. arXiv preprint arXiv:2504.02821, 2025

    arXiv 2025

  52. [52]

    A framework to learn with interpretation

    Jayneel Parekh, Pavlo Mozharovskyi, and Florence d'Alch \'e - Buc. A framework to learn with interpretation. Advances in Neural Information Processing Systems, 34: 0 24273--24285, 2021

    2021

  53. [53]

    Listen to interpret: Post-hoc interpretability for audio networks with nmf

    Jayneel Parekh, Sanjeel Parekh, Pavlo Mozharovskyi, Florence d'Alch \'e - Buc, and Ga \"e l Richard. Listen to interpret: Post-hoc interpretability for audio networks with nmf. Advances in Neural Information Processing Systems, 35: 0 35270--35283, 2022

    2022

  54. [54]

    Restyling unsupervised concept based interpretable networks with generative models

    Jayneel Parekh, Quentin Bouniot, Pavlo Mozharovskyi, Alasdair Newson, and Florence d'Alch \'e Buc. Restyling unsupervised concept based interpretable networks with generative models. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=CexatBp6rx

    2025

  55. [55]

    On estimation of a probability density function and mode

    Emanuel Parzen. On estimation of a probability density function and mode. The Annals of Mathematical Statistics, 33 0 (3): 0 1065--1076, 1962

    1962

  56. [56]

    Causality

    Judea Pearl. Causality. Cambridge university press, 2009

    2009

  57. [57]

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

    Owen Queen, Thomas Hartvigsen, Teddy Koker, Huan He, Theodoros Tsiligkaridis, and Marinka Zitnik. Encoding time-series explanations through self-supervised model behavior consistency. In NeurIPS, 2023

    2023

  58. [58]

    Improving sparse decomposition of language model activations with gated sparse autoencoders

    Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum, Vikrant Varma, Janos Kramar, Rohin Shah, and Neel Nanda. Improving sparse decomposition of language model activations with gated sparse autoencoders. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024 a

    2024

  59. [59]

    Jumping ahead: Improving reconstruction fidelity with jumprelu sparse autoencoders

    Senthooran Rajamanoharan, Tom Lieberum, Nicolas Sonnerat, Arthur Conmy, Vikrant Varma, J \'a nos Kram \'a r, and Neel Nanda. Jumping ahead: Improving reconstruction fidelity with jumprelu sparse autoencoders. arXiv preprint arXiv:2407.14435, 2024 b

    Pith/arXiv arXiv 2024

  60. [60]

    Introducing a new benchmarked dataset for activity monitoring

    Attila Reiss and Didier Stricker. Introducing a new benchmarked dataset for activity monitoring. In ISWC, pp.\ 108--109, 2012

    2012

  61. [61]

    `` W hy should I trust you?" E xplaining the predictions of any classifier

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. `` W hy should I trust you?" E xplaining the predictions of any classifier. In SIGKDD, pp.\ 1135--1144, 2016

    2016

  62. [62]

    Time series of heat demand and heat pump efficiency for energy system modeling

    Oliver Ruhnau, Lion Hirth, and Aaron Praktiknjo. Time series of heat demand and heat pump efficiency for energy system modeling. Scientific data, 6 0 (1): 0 1--10, 2019

    2019

  63. [63]

    A framework for learning ante-hoc explainable models via concepts

    Anirban Sarkar, Deepak Vijaykeerthy, Anindya Sarkar, and Vineeth N Balasubramanian. A framework for learning ante-hoc explainable models via concepts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10286--10295, 2022

    2022

  64. [64]

    The power of explainability in forecast-informed deep learning models for flood mitigation

    Jimeng Shi, Vitalii Stebliankin, and Giri Narasimhan. The power of explainability in forecast-informed deep learning models for flood mitigation. arXiv preprint arXiv:2310.19166, 2023

    Pith/arXiv arXiv 2023

  65. [65]

    Axiomatic attribution for deep networks

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In ICML, pp.\ 3319--3328, 2017

    2017

  66. [66]

    Clinical intervention prediction and understanding with deep neural networks

    Harini Suresh, Nathan Hunt, Alistair Johnson, Leo Anthony Celi, Peter Szolovits, and Marzyeh Ghassemi. Clinical intervention prediction and understanding with deep neural networks. In MLHC, pp.\ 322--337, 2017

    2017

  67. [67]

    What went wrong and when? I nstance-wise feature importance for time-series black-box models

    Sana Tonekaboni, Shalmali Joshi, Kieran Campbell, David K Duvenaud, and Anna Goldenberg. What went wrong and when? I nstance-wise feature importance for time-series black-box models. In NeurIPS, pp.\ 799--809, 2020

    2020

  68. [68]

    Start smart: Leveraging gradients for enhancing mask-based xai methods

    Buelent Uendes, Shujian Yu, and Mark Hoogendoorn. Start smart: Leveraging gradients for enhancing mask-based xai methods. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  69. [69]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, pp.\ 5998--6008, 2017

    2017

  70. [70]

    Counterfactual explanations for time series forecasting

    Zhendong Wang, Ioanna Miliou, Isak Samsten, and Panagiotis Papapetrou. Counterfactual explanations for time series forecasting. In 2023 IEEE International Conference on Data Mining (ICDM), pp.\ 1391--1396. IEEE, 2023

    2023

  71. [71]

    Provable compositional generalization for object-centric learning

    Thadd \"a us Wiedemer, Jack Brady, Alexander Panfilov, Attila Juhos, Matthias Bethge, and Wieland Brendel. Provable compositional generalization for object-centric learning. In The Twelfth International Conference on Learning Representations, 2024

    2024

  72. [72]

    Vitpose: Simple vision transformer baselines for human pose estimation

    Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in neural information processing systems, 35: 0 38571--38584, 2022

    2022

  73. [73]

    Self-interpretable time series prediction with counterfactual explanations

    Jingquan Yan and Hao Wang. Self-interpretable time series prediction with counterfactual explanations. In International Conference on Machine Learning, pp.\ 39110--39125. PMLR, 2023

    2023

  74. [74]

    Explain temporal black-box models via functional decomposition

    Linxiao Yang, Yunze Tong, Xinyue Gu, and Liang Sun. Explain temporal black-box models via functional decomposition. In Forty-first International Conference on Machine Learning, 2024

    2024

  75. [75]

    Time series shapelets: a new primitive for data mining

    Lexiang Ye and Eamonn Keogh. Time series shapelets: a new primitive for data mining. In Proceedings of the 15th ACM SIGKDD international conference on Knowledge discovery and data mining, pp.\ 947--956, 2009

    2009

  76. [76]

    Optimal information retention for time-series explanations

    Jinghang Yue, Jing Wang, Lu Zhang, Shuo Zhang, Da Li, Zhaoyang Ma, and Youfang Lin. Optimal information retention for time-series explanations. In Forty-second International Conference on Machine Learning, 2025

    2025

  77. [77]

    Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factors

    Zeyu Yun, Yubei Chen, Bruno A Olshausen, and Yann LeCun. Transformer visualization via dictionary learning: contextualized embedding as a linear superposition of transformer factors. arXiv preprint arXiv:2103.15949, 2021

    Pith/arXiv arXiv 2021

  78. [78]

    Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pp.\ 11121--11128, 2023

    Ailing Zeng, Muxi Chen, Lei Zhang, and Qiang Xu. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence, volume 37, pp.\ 11121--11128, 2023

    2023

  79. [79]

    Informer: Beyond efficient transformer for long sequence time-series forecasting

    Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang. Informer: Beyond efficient transformer for long sequence time-series forecasting. In Proceedings of the AAAI conference on artificial intelligence, pp.\ 11106--11115, 2021

    2021

  80. [80]

    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...