Pith. sign in

REVIEW 3 major objections 6 minor 96 references

IB-Forecast matches leading black-box multivariate forecasting error while emitting sparse explanations that are faithful by construction, at no extra inference cost.

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 · grok-4.5

2026-07-31 17:08 UTC pith:C2GGHF6W

load-bearing objection Solid engineering paper: residual IB gates give near-SOTA multivariate forecasts with sparse native masks that beat post-hoc and other native explainers under matched-budget fidelity, with the main caveat that faithfulness is partly by construction. the 3 major comments →

arxiv 2607.28124 v1 pith:C2GGHF6W submitted 2026-07-30 cs.LG cs.AI

Information Bottleneck Learning for Faithful Time Series Forecasting Explanations

Xu Zheng , Wei Cheng , Zhuomin Chen , Mo Sha , Jingchao Ni , Dongsheng Luo This is my paper
classification cs.LG cs.AI
keywords time series forecastinginterpretabilityfaithfulnessinformation bottlenecksparse masksmultivariate forecastingseasonal decomposition
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.

When forecasts drive real decisions, knowing which past observations the model actually used matters as much as the numbers it outputs. Existing “interpretable” forecasters expose internal pieces but rarely check whether those pieces truly determine the prediction, and faithfulness tools from classification do not transfer cleanly to multi-step forecasting. This paper argues that the gap can be closed by design: separate each series into a disclosed structural context (level, scale, and a learned seasonal profile) and a residual path, then force all window-specific dependence through a budgeted binary mask over residual tokens. Trained end-to-end with an information-bottleneck objective, the model jointly produces the forecast and the mask in one forward pass. Across standard multivariate benchmarks it stays competitive with strong black-box models while needing only about 14–20% of the history, and under matched sparsity its native masks beat gradient, occlusion, and test-time optimization baselines on faithfulness.

Core claim

A multivariate forecaster can be made self-interpretable without an accuracy tax by decomposing each prediction into a disclosed periodic-and-frame context plus a gated residual readout, then training a sparse binary mask as an information bottleneck so that, conditioned on that context, the forecast depends only on the selected residual tokens. Under a matched sparsity budget those native masks achieve higher matched-budget fidelity than post-hoc and other native explainers, while reading only 14–20% of the observations.

What carries the argument

Budget-constrained information bottleneck over residual patch–channel tokens: after instance normalization and subtraction of a learned seasonal profile Q, a gate network emits hard binary masks; the readout sees only masked deviations, so closed tokens are structurally prevented from affecting the forecast once context is fixed.

Load-bearing premise

Once level, scale, and a fixed-period seasonal profile are treated as known context, the useful window-specific signal lives in a small set of residual tokens, and “deleting” a token is correctly modeled by resetting it to that seasonal baseline.

What would settle it

On held-out windows, fix the disclosed context and the model’s own open-token count: if keeping only the native mask’s tokens fails to reproduce the forecast (high sufficiency error) or removing them fails to change it (low comprehensiveness) relative to strong post-hoc masks—or if forcing a 14–20% read rate sharply raises error on series whose drivers are not residual to a fixed cycle—the central claim fails.

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

If this is right

  • Users can dial explanation sparsity with a single budget and still obtain competitive multi-horizon forecasts.
  • Faithfulness can be architecturally guaranteed for forecasting rather than checked only after the fact on frozen models.
  • Existing native structures (attention weights, patch contributions, linear maps) can score poorly on matched-budget fidelity even when accuracy looks fine.
  • Accurate long-horizon forecasts on common benchmarks often need only a small fraction of the look-back once seasonality and scale are disclosed.
  • Explanation cost collapses to zero at inference because the mask is emitted in the same forward pass as the forecast.

Where Pith is reading between the lines

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

  • The same residual-plus-budget pattern may transfer to other structured prediction settings where a stable baseline can be disclosed and only deviations must be attributed.
  • If level or scale shifts themselves carry the decision-relevant signal, the method would need an explicit second bottleneck over the frame, not only over residuals.
  • Matched-budget fidelity with in-distribution seasonal imputation could become a default reporting standard for any claim of “interpretable forecasting.”
  • Tight budgets may act as a regularizer that discards brittle correlations, which is a testable claim on distribution-shift splits.

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

3 major / 6 minor

Summary. IB-Forecast is a self-interpretable multivariate forecasting architecture that, in one forward pass, emits both a multi-step forecast and a sparse explanation. It discloses a structural context C (phase, per-channel level/scale, and a learned seasonal profile Q), maps the window to residual patch×channel tokens, and routes all window-specific dependence through a budgeted hard binary mask trained with an information-bottleneck-style objective. The forecast is reconstructed as seasonal profile plus a gated residual readout, so closed tokens are architecturally prevented from affecting the residual path (Problem 1; Eqs. 2, 7–8). Empirically, on standard LTSF benchmarks the model matches strong black-box forecasters (TQNet/CycleNet) in MSE/MAE while using only ~14–20% of tokens; under matched-budget comprehensiveness/sufficiency its native masks beat gradient, occlusion, test-time mask optimizers, and the native attributions of TFT/PatchDecomp/DLinear. A synthetic planted-support benchmark and ablations (gate, budget, cycle Q) support the design.

Significance. The paper targets a real and under-served gap: interpretable-by-design forecasters rarely measure faithfulness, while faithfulness-oriented IB/mask methods are almost entirely post-hoc and classification-oriented. Formulating self-interpretable multi-step forecasting with disclosed structural context, amortized hard gates, and a matched-budget fidelity protocol adapted to seasonal-residual semantics is a substantive contribution. Strengths include end-to-end sparsity control, zero extra inference cost for explanations, explicit architectural faithfulness (Rudin-style), a clear accuracy–sparsity frontier (Figure 2), synthetic ground-truth recovery (Table 8), and appendix theory linking relevance to an error floor. If the residual-mask-plus-seasonal-context assumption holds on the target domains, the work offers a practical path to certified sparse explanations without sacrificing competitive LTSF accuracy.

major comments (3)
  1. [Problem 1; Table 2; Faithfulness under matched-budget fidelity] Table 2 vs Problem 1 / Eqs. 7–8: Conditional faithfulness is explicitly by construction (Ŷ=g(E;C); closed tokens zeroed in deviation space). Consequently, very high Comp and very low Suff for the native mask are partly definitional once the evaluation freezes (μ,σ) and imputes deleted tokens with the same seasonal reference the architecture uses. The paper should more sharply separate three claims: (i) architectural guarantee, (ii) empirical superiority of IB-Forecast’s native masks over other models’ native explanations (TFT/PatchDecomp negative Scores are the strongest non-circular evidence), and (iii) synthetic support recovery (Table 8). As written, the abstract’s phrasing that native explanations “consistently surpass” post-hoc methods applied to the same model overstates an advantage that the architecture partly installs. A short dedicated paragraph in the faithfulness section woul
  2. [Abstract; Table 1; Table 7 (Breadth on large-channel datasets)] Table 1 / abstract vs Table 7: The headline claim that IB-Forecast “matches the forecasting error of leading black-box models” is supported on the five main datasets (overall MSE tied with TQNet), but on Traffic (C=862) the gap is about +9% average MSE, and Electricity/Solar show smaller but nonzero gaps. The abstract and introduction should qualify the parity claim by regime (moderate C, clear periodicity) or move large-C numbers into the main comparison. Otherwise readers will take “no interpretability tax” as universal, which Table 7 does not support.
  3. [Training objective Eq. (9); Theoretical Analysis Eq. (10); Theorem 1] Eq. (9) vs Eq. (10) / Theorem 1: The training loss uses a Bernoulli-KL gate regularizer, a quadratic budget penalty, and TV—not a direct estimator of I(X;E|C) or I(E;Y|C). Theorem 1’s error-floor bound is standard and useful, but the manuscript presents the objective as implementing the conditional IB. Please state explicitly that Eq. (9) is a tractable surrogate (in the GSAT/L0 tradition) and that Theorem 1 bounds any readout of the form g(E;C) rather than certifying that the learned gates achieve the IB optimum. This is a framing fix, not a derivation error.
minor comments (6)
  1. [Related Work; Experiments / Baselines] N-BEATS and N-HiTS are discussed in Related Work as canonical interpretable-by-design forecasters but are absent from the experimental tables. A one-sentence justification (e.g., basis decompositions are not input-attribution explanations under the matched-budget protocol) would help.
  2. [Table 1; Abstract] Table 1 omits Electricity from the main average while the abstract invokes energy applications; either include it in the main table or note that large-C results are deferred to the appendix.
  3. [Figure 3] Figure 3 case study is helpful but hard to read in grayscale; consider marking open gates with hatching or higher-contrast outlines in addition to green shading.
  4. [Table 1 caption; Baseline details] ProtoTS Weather cells are omitted for compute budget; state the hardware and wall-clock cutoff once in the main text so the omission is reproducible.
  5. [Methodology; Preliminaries] Minor notation: D and Δ are used interchangeably for the residual after Eq. (2); pick one symbol in the main text.
  6. [Table 2] The TIMING row is oracle-tuned over segment settings (Appendix); mark this more visibly in Table 2’s header so readers do not treat it as a default-hyperparameter baseline.

Circularity Check

2 steps flagged

Native-mask fidelity is largely definitional: architecture forces Ŷ=g(E;C) and the fidelity protocol reuses the same absent-input semantics, while accuracy and cross-model comparisons remain independent.

specific steps
  1. self definitional [Problem 1; Methodology (Gated deviation readout); Abstract]
    "Conditional faithfulness by construction: conditioned on C, the forecast depends only on E; that is, Ŷ=g(E;C) with the readout function g. ... Crucially, IB-Forecast guarantees faithfulness by architecturally ensuring that only the tokens selected by the information bottleneck can affect the deviation readout. ... When a token is closed, it contributes zero in deviation space."

    The paper’s guarantee that the emitted mask is a faithful explanation is identical to the architectural definition of the forecaster: beyond disclosed C, Ŷ is computed only from m⊙D. High native-mask faithfulness is therefore not an independent derived property; it is the construction restated. Sufficiency of the true open set is forced toward zero whenever hard gates and the gated readout are implemented as specified.

  2. self definitional [Methodology (Gated deviation readout); Experiments (Metrics)]
    "The same interpretation is used in our evaluation of faithfulness. Deleting a token resets it to the seasonal profile while keeping the structural context fixed, so the architecture and evaluation metric share the same definition of an absent input. ... (iii) we impute deleted values using the model’s own reference semantics (μ+σ Q(t) or the window mean)."

    Matched-budget Comp/Suff for IB-Forecast’s native mask reuse the model’s own closed-token semantics (reset to μ+σ⊙Q, freeze RevIN stats). Architecture and metric are co-defined, so superior native Score vs post-hoc methods on this model partly reflects measuring the installed computation against itself, not an external faithfulness criterion. Cross-model native comparisons and accuracy-under-budget results are not dissolved by this alignment.

full rationale

IB-Forecast does not hide a sneaky fit-as-prediction loop. It explicitly designs conditional faithfulness into the forward pass (Problem 1: Ŷ=g(E;C); hard gates zero closed deviation tokens) and then scores explanations with a matched-budget protocol that deletes inputs by resetting them to the same seasonal baseline μ+σ⊙Q the model assumes. For the native mask on its own hard-gated model, very low sufficiency and high comprehensiveness are therefore largely forced by construction rather than independently discovered. That is real but partial circularity on the “guarantees high explanation fidelity” slogan. The load-bearing empirical content is not thereby emptied: matching black-box MSE/MAE under a tight ρ budget, the ρ–error frontier, the cycle-Q and hard-gate ablations, synthetic planted-support recovery, and especially the failure of TFT/PatchDecomp native maps on their own forecasts are independent checks that do not reduce to the definition. No load-bearing self-citation uniqueness theorem or fitted-parameter-as-prediction chain was found. Score 4 reflects definitional alignment of architecture and metric on one central claim, with substantial non-circular empirical remainder.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

The load-bearing content is an engineering method plus empirical protocol. It rests on classical additive decomposition and IB ideas, plus modeling choices that residuals after RevIN and a fixed-period learned profile leave a sparse token mask as sufficient window-specific evidence. Free parameters are the usual deep-learning knobs (budget, IB weight, gate priors, widths). No new physical entity is postulated; the ‘explanation object’ is an architectural construct.

free parameters (5)
  • sparsity budget ρ = 0.2 (main fidelity); swept 0.05–0.5
    User-chosen open-rate target that defines the operating point of Problem 1 and the matched-budget comparisons; main claims use ~0.2 (realized 14–20%).
  • IB/KL weight β = ~0.02 real data; 0.1 synthetic
    Trades forecast loss against gate sparsity; set per dataset/config and shown to control decoy rejection on synthetic data.
  • budget penalty λb and TV weight λtv = λb≈5, λtv≈1e-3 (defaults in Fig. 5)
    Extra loss coefficients enforcing mean open rate and temporal contiguity of masks; defaults given in sensitivity sweeps.
  • Bernoulli prior π and gate temperature T = π=0.5 training table / 0.2 in sweeps; T=0.5 annealed
    Prior open probability and Concrete/STE temperature schedule for stochastic hard gates.
  • architecture knobs (patch length, d, gate layers, readout type, W) = See Table 5; W=24/96/144 by dataset
    Patch length 12, width 64, 1–2 gate layers, linear vs MLP readout, cycle length W from sampling rate—chosen/validated per dataset.
axioms (5)
  • domain assumption Additive classical decomposition into trend/seasonal/residual and invertible RevIN-style normalization once (μ,σ,Q,τ) are fixed.
    Preliminaries Eq. (1)–(2) and assumption (A2); underpins treating closed tokens as seasonal baseline.
  • domain assumption Disclosed structural context C spans both look-back and horizon by periodic extension of Q.
    Stated in Preliminaries and used in Eq. (8); required for Ŷ=g(E;C).
  • domain assumption Conditional information bottleneck min −I(E;Y|C)+β I(X;E|C) is an appropriate training principle for sparse faithful explanations.
    Methodology and Theorem 3; imported from Tishby IB and GSAT-style stochastic gates.
  • standard math Straight-through hard-Concrete / binary gates yield a valid trainable approximation to discrete masks.
    Eq. (6); standard STE/Concrete practice (Bengio, Maddison, Louizos).
  • ad hoc to paper Matched-budget comprehensiveness/sufficiency with frozen (μ,σ) and seasonal imputation is the right faithfulness metric for multi-step forecasts.
    Metrics section adaptations of ERASER; protocol choices materially shape Table 2 rankings.
invented entities (2)
  • IB-Forecast explanation object E=(mask m, masked deviations) under disclosed C independent evidence
    purpose: Single forward-pass compact attribution that is also the residual computational pathway.
    Defined in Problem 1 and Algorithm 1; not an external physical entity, but the paper’s core constructed object.
  • Amortized transformer gate network over patch×channel residual tokens no independent evidence
    purpose: Produce instance-wise sparse binary masks without test-time optimization.
    Stages 2–3; empirical entity validated by fidelity and synthetic AUROC, not independently measured outside this architecture.

pith-pipeline@v1.2.0-daily-grok45 · 30252 in / 3729 out tokens · 83750 ms · 2026-07-31T17:08:26.720716+00:00 · methodology

0 comments
read the original abstract

As forecasts increasingly drive decisions in fields such as energy, transportation, and healthcare, understanding the historical data behind these predictions has become as crucial as the predictions themselves. Although existing interpretable-by-design forecasters reveal their internal structures, they offer no guarantee that these structures faithfully reflect the underlying evidence driving the predictions. In contrast, while faithfulness-oriented methods explicitly verify model behavior, they are almost exclusively designed for post-hoc classification tasks. To bridge this gap, we propose IB-Forecast, an inherently interpretable multivariate time-series forecasting framework. It decomposes forecasting into a learned periodic component and a residual component computed with explainable masks over input tokens. With a budget-constrained information bottleneck, end-to-end optimization enables users to directly control explanation sparsity. With a rigorous faithfulness evaluation protocol, extensive experiments demonstrate that IB-Forecast matches the forecasting error of leading black-box models while providing faithful explanations at no additional inference cost. Furthermore, under a matched sparsity budget, these native explanations consistently surpass gradient-based, occlusion-based, and optimization-based baselines across all evaluated datasets. Ultimately, whereas the native explanations of existing interpretable forecasters exhibit poor faithfulness, IB-Forecast guarantees high explanation fidelity, requiring only 14-20% of the observations to deliver low-error predictions.

Figures

Figures reproduced from arXiv: 2607.28124 by Dongsheng Luo, Jingchao Ni, Mo Sha, Wei Cheng, Xu Zheng, Zhuomin Chen.

Figure 1
Figure 1. Figure 1: Overview of IB-Forecast. Stage 1 normalizes the input [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Trade-off between explanation sparsity and forecasting error. Test MSE (left) and MAE (right) against the measured [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: An example of IB-Forecast explanation for one [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Gate-open frequency per (lag-patch, channel) over [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: One-at-a-time hyperparameter sensitivity on ETTh1, ETTm1 and Weather ( [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Fidelity against inference cost per explanation on [PITH_FULL_IMAGE:figures/full_fig_p017_6.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

96 extracted references · 4 canonical work pages

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    1984 , issn =

    Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

  9. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  10. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  11. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  12. [12]

    2024 , month =

    Liu, Zichuan and Wang, Tianchun and Shi, Jimeng and Zheng, Xu and Chen, Zhuomin and Song, Lei and Dong, Wenqian and Obeysekera, Jayantha and Shirani, Farhad and Luo, Dongsheng , booktitle =. 2024 , month =. 2405.09308 , archivePrefix =

    Pith/arXiv arXiv 2024

  13. [13]

    ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting , year =

    Ziheng Peng and Shijie Ren and Xinyue Gu and Linxiao Yang and Xiting Wang and Liang Sun , booktitle =. ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting , year =

  14. [14]

    2025 , eprint=

    A Self-explainable Model of Long Time Series by Extracting Informative Structured Causal Patterns , author=. 2025 , eprint=

    2025

  15. [15]

    2026 , eprint=

    When, How Long and How Much? Interpretable Neural Networks for Time Series Regression by Learning to Mask and Aggregate , author=. 2026 , eprint=

    2026

  16. [16]

    2024 , eprint=

    TimeSieve: Extracting Temporal Dynamics through Information Bottlenecks , author=. 2024 , eprint=

    2024

  17. [17]

    Towards Generalizable Time Series Forecasting Via IB-Regularized Transformer-Diffusion , year=

    Na, Dagyeong and Kang, Jinho and Kwon, Junseok , booktitle=. Towards Generalizable Time Series Forecasting Via IB-Regularized Transformer-Diffusion , year=

  18. [18]

    TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop , url =

    Jiang, Yushan and Yu, Wenchao and Lee, Geon and Song, Dongjin and Shin, Kijung and Cheng, Wei and Liu, Yanchi and Chen, Haifeng , booktitle =. TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop , url =

  19. [19]

    Proceedings of the 42nd International Conference on Machine Learning , pages =

    Optimal Information Retention for Time-Series Explanations , author =. Proceedings of the 42nd International Conference on Machine Learning , pages =. 2025 , editor =

    2025

  20. [20]

    The Fourteenth International Conference on Learning Representations , year=

    TimeSeg: An Information-Theoretic Segment-Wise Explainer for Time-Series Predictions , author=. The Fourteenth International Conference on Learning Representations , year=

  21. [21]

    Contimask: Explaining Irregular Time Series via Perturbations in Continuous Time , url =

    Moebus, Max and Braun, Bj\". Contimask: Explaining Irregular Time Series via Perturbations in Continuous Time , url =. Advances in Neural Information Processing Systems , editor =

  22. [22]

    World Conference on Explainable Artificial Intelligence , pages=

    FLEXtime: Filterbank learning to explain time series , author=. World Conference on Explainable Artificial Intelligence , pages=. 2025 , organization=

    2025

  23. [23]

    Start Smart: Leveraging Gradients For Enhancing Mask-based

    Buelent Uendes and Shujian Yu and Mark Hoogendoorn , booktitle=. Start Smart: Leveraging Gradients For Enhancing Mask-based. 2025 , url=

    2025

  24. [24]

    2025 , isbn =

    Zhang, Shuo and Wang, Jing and Nie, Shiqin and Yue, Jinghang and Zhu, Weikang and Lin, Youfang , title =. 2025 , isbn =. doi:10.1145/3711896.3737033 , booktitle =

    arXiv 2025

  25. [25]

    Proceedings of the 10th Machine Learning for Healthcare Conference , year =

    Monte Carlo ExtremalMask: Uncertainty Aware Time Series Model Interpretability For Critical Care Applications , author =. Proceedings of the 10th Machine Learning for Healthcare Conference , year =

  26. [26]

    2025 , month =

    Jang, Hyeongwon and Kim, Changhun and Yang, Eunho , booktitle =. 2025 , month =

    2025

  27. [27]

    Forty-third International Conference on Machine Learning , year=

    Unified Time Series Explanations via Amortized Optimization and Instance-level Multi-Expert Knowledge Distillation , author=. Forty-third International Conference on Machine Learning , year=

  28. [28]

    , title =

    Tran, Viet-Hung and Zhang, Zichi and Pham, Tuan Dung and Doan, Ngoc Phu and Hoang, Anh-Tuan and Li, Peixin and Vandierendonck, Hans and Assent, Ira and Mai, Son T. , title =. doi:10.1609/aaai.v39i20.35387 , booktitle =

    work page doi:10.1609/aaai.v39i20.35387

  29. [29]

    ShapeX: Shapelet-Driven Post Hoc Explanations for Time Series Classification Models , url =

    Huang, Bosong and Jin, Ming and Liang, Yuxuan and Barthelemy, Johan and Cheng, Debo and Wen, Qingsong and Liu, Chenghao and Pan, Shirui , booktitle =. ShapeX: Shapelet-Driven Post Hoc Explanations for Time Series Classification Models , url =

  30. [30]

    2025 , eprint=

    DeltaSHAP: Explaining Prediction Evolutions in Online Patient Monitoring with Shapley Values , author=. 2025 , eprint=

    2025

  31. [31]

    Changhun Kim and Yechan Mun and Hyeongwon Jang and Eunseo Lee and Sangchul Hahn and Eunho Yang , booktitle=. Delta-. 2026 , url=

    2026

  32. [32]

    Smith , year=

    Fuyuan Lyu and Wesley Chung and Melissa Mozifian and Ankit Vani and Edward J. Smith , year=

  33. [33]

    Sensors , VOLUME =

    Wang, Jiazhan and Zhang, Ruifeng and Li, Qiang , TITLE =. Sensors , VOLUME =. 2025 , NUMBER =

    2025

  34. [34]

    Greeks in AI Symposium 2025 , year=

    Time series saliency maps: Explaining models across multiple domains , author=. Greeks in AI Symposium 2025 , year=

    2025

  35. [35]

    2026 , issue_date =

    Park, Youngjin and Tong, Anh and Lee, Sehyun and Seong, Jihyeon and Xie, Qin and Choi, Jaesik , title =. 2026 , issue_date =. doi:10.1145/3794839 , journal =

    work page doi:10.1145/3794839 2026

  36. [36]

    TechRxiv , volume =

    Zhuomin Chen and Gabriel Lucchesi and Qingkai Dong and Xu Zheng and Dongjin Song and Qingsong Wen and Wei Cheng and Jingchao Ni and Dongsheng Luo , title =. TechRxiv , volume =. 2026 , doi =

    2026

  37. [37]

    Proceedings of the AAAI Symposium Series , volume=

    Temporal Concept Tracing: Making Deep Learning Predictions Interpretable and Actionable for ICU Acute Kidney Injury Prevention , author=. Proceedings of the AAAI Symposium Series , volume=

  38. [38]

    Time- RA : Towards Time Series Reasoning for Anomaly Diagnosis with LLM Feedback

    Yang, Yiyuan and Liu, Zichuan and Song, Lei and Ying, Kai and Wang, Stephen and Bamford, Joshua Thomas and Vyetrenko, Svitlana and Bian, Jiang and Wen, Qingsong. Time- RA : Towards Time Series Reasoning for Anomaly Diagnosis with LLM Feedback. Findings of the A ssociation for C omputational L inguistics: ACL 2026. 2026. doi:10.18653/v1/2026.findings-acl.562

    work page doi:10.18653/v1/2026.findings-acl.562 2026

  39. [39]

    Khalid Oublal and Quentin Bouniot and Qi Gan and Stephan Cl. Time. Forty-third International Conference on Machine Learning , year=

  40. [40]

    Encoding Time-Series Explanations through Self-Supervised Model Behavior Consistency , url =

    Queen, Owen and Hartvigsen, Tom and Koker, Teddy and He, Huan and Tsiligkaridis, Theodoros and Zitnik, Marinka , booktitle =. Encoding Time-Series Explanations through Self-Supervised Model Behavior Consistency , url =

  41. [41]

    International Conference on Learning Representations , year=

    Explaining Time Series via Contrastive and Locally Sparse Perturbations , author=. International Conference on Learning Representations , year=

  42. [42]

    2025 , eprint=

    Counterfactual Explanation for Multivariate Time Series Forecasting with Exogenous Variables , author=. 2025 , eprint=

    2025

  43. [43]

    International Conference on Learning Representations , year=

    N-BEATS: Neural basis expansion analysis for interpretable time series forecasting , author=. International Conference on Learning Representations , year=

  44. [44]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Nhits: Neural hierarchical interpolation for time series forecasting , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  45. [45]

    2021 , doi =

    Temporal Fusion Transformers for interpretable multi-horizon time series forecasting , journal =. 2021 , doi =

    2021

  46. [46]

    2025 , eprint=

    Interpretability for Time Series Transformers using A Concept Bottleneck Framework , author=. 2025 , eprint=

    2025

  47. [47]

    2026 , eprint=

    OceanCBM: A Concept Bottleneck Model for Mechanistic Interpretability in Ocean Forecasting , author=. 2026 , eprint=

    2026

  48. [48]

    2025 , eprint=

    Signals, Concepts, and Laws: Toward Universal, Explainable Time-Series Forecasting , author=. 2025 , eprint=

    2025

  49. [49]

    2025 , eprint=

    iTFKAN: Interpretable Time Series Forecasting with Kolmogorov-Arnold Network , author=. 2025 , eprint=

    2025

  50. [50]

    2026 , eprint=

    PatchDecomp: Interpretable Patch-Based Time Series Forecasting , author=. 2026 , eprint=

    2026

  51. [51]

    FocusLearn: Fully-Interpretable, High-Performance Modular Neural Networks for Time Series , year=

    Su, Qiqi and Kloukinas, Christos and d’Avila Garcez, Artur , booktitle=. FocusLearn: Fully-Interpretable, High-Performance Modular Neural Networks for Time Series , year=

  52. [52]

    2024 , eprint=

    EDformer: Embedded Decomposition Transformer for Interpretable Multivariate Time Series Predictions , author=. 2024 , eprint=

    2024

  53. [53]

    2026 , eprint=

    Towards Accurate and Interpretable Time-series Forecasting: A Polynomial Learning Approach , author=. 2026 , eprint=

    2026

  54. [54]

    2026 , doi =

    WEITS: A wavelet-enhanced residual framework for interpretable time series forecasting , journal =. 2026 , doi =

    2026

  55. [55]

    2024 , eprint=

    Interpretable Multivariate Time Series Forecasting Using Neural Fourier Transform , author=. 2024 , eprint=

    2024

  56. [56]

    2026 , eprint=

    Dual-Prototype Disentanglement: A Context-Aware Enhancement Framework for Time Series Forecasting , author=. 2026 , eprint=

    2026

  57. [57]

    Data Mining and Knowledge Discovery , volume=

    Prototsnet: interpretable multivariate time series classification with prototypical parts , author=. Data Mining and Knowledge Discovery , volume=. 2026 , publisher=

    2026

  58. [58]

    2019 , publisher =

    Ming, Yao and Xu, Panpan and Qu, Huamin and Ren, Liu , title =. 2019 , publisher =. doi:10.1145/3292500.3330908 , booktitle =

    arXiv 2019

  59. [59]

    Counterfactual Explanations for Time Series Forecasting , year=

    Wang, Zhendong and Miliou, Ioanna and Samsten, Isak and Papapetrou, Panagiotis , booktitle=. Counterfactual Explanations for Time Series Forecasting , year=

  60. [60]

    Navigating Time’s Possibilities: Plausible Counterfactual Explanations for Multivariate Time-Series Forecast through Genetic Algorithms , year=

    Zuin, Gianlucca and Veloso, Adriano , booktitle=. Navigating Time’s Possibilities: Plausible Counterfactual Explanations for Multivariate Time-Series Forecast through Genetic Algorithms , year=

  61. [61]

    Instance-Based Counterfactual Explanations for Time Series Classification

    Delaney, Eoin and Greene, Derek and Keane, Mark T. Instance-Based Counterfactual Explanations for Time Series Classification. Case-Based Reasoning Research and Development. 2021

    2021

  62. [62]

    2026 , eprint=

    Revisiting Multivariate Time Series Forecasting with Missing Values , author=. 2026 , eprint=

    2026

  63. [63]

    Forty-third International Conference on Machine Learning , year=

    What If We Let Forecasting Forget? A Sparse Bottleneck for Cross-Variable Dependencies , author=. Forty-third International Conference on Machine Learning , year=

  64. [64]

    Latent Representation and Simulation of Markov Processes via Time-Lagged Information Bottleneck , url =

    Federici, Marco and Forr\'. Latent Representation and Simulation of Markov Processes via Time-Lagged Information Bottleneck , url =. The Twelfth International Conference on Learning Representations , editor =

  65. [65]

    On the Necessity of Multi-Domain Explanation: An Uncertainty Principle Approach for Deep Time Series Models , year=

    Rezaei, Shahbaz and Halev, Avishai and Liu, Xin , booktitle=. On the Necessity of Multi-Domain Explanation: An Uncertainty Principle Approach for Deep Time Series Models , year=

  66. [66]

    F-Fidelity: A Robust Framework for Faithfulness Evaluation of Explainable AI , url =

    Zheng, Xu and Shirani, Farhad and Chen, Zhuomin and Lin, Chaohao and Cheng, Wei and Guo, Wenbo and Luo, Dongsheng , booktitle =. F-Fidelity: A Robust Framework for Faithfulness Evaluation of Explainable AI , url =

  67. [67]

    Proceedings of the 39th International Conference on Machine Learning , pages =

    Interpretable and Generalizable Graph Learning via Stochastic Attention Mechanism , author =. Proceedings of the 39th International Conference on Machine Learning , pages =. 2022 , month =

    2022

  68. [68]

    International Conference on Machine Learning , pages =

    Temporal Query Network for Efficient Multivariate Time Series Forecasting , author =. International Conference on Machine Learning , pages =. 2025 , volume =

    2025

  69. [69]

    CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns , url =

    Lin, Shengsheng and Lin, Weiwei and Hu, Xinyi and Wu, Wentai and Mo, Ruichao and Zhong, Haocheng , booktitle =. CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns , url =. doi:10.52202/079017-3373 , pages =

    work page doi:10.52202/079017-3373

  70. [70]

    TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables , url =

    Wang, Yuxuan and Wu, Haixu and Dong, Jiaxiang and Qin, Guo and Zhang, Haoran and Liu, Yong and Qiu, Yunzhong and Wang, Jianmin and Long, Mingsheng , booktitle =. TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables , url =. doi:10.52202/079017-0015 , editor =

    doi:10.52202/079017-0015

  71. [71]

    iTransformer: Inverted Transformers Are Effective for Time Series Forecasting , url =

    Liu, Yong and Hu, Tengge and Zhang, Haoran and Wu, Haixu and Wang, Shiyu and Ma, Lintao and Long, Mingsheng , booktitle =. iTransformer: Inverted Transformers Are Effective for Time Series Forecasting , url =

  72. [72]

    International Conference on Learning Representations , year=

    A Time Series is Worth 64 Words: Long-term Forecasting with Transformers , author=. International Conference on Learning Representations , year=

  73. [73]

    International Conference on Learning Representations , year=

    TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis , author=. International Conference on Learning Representations , year=

  74. [74]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Are transformers effective for time series forecasting? , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  75. [75]

    International Conference on Learning Representations , year=

    Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift , author=. International Conference on Learning Representations , year=

  76. [76]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Informer: Beyond efficient transformer for long sequence time-series forecasting , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  77. [77]

    International Conference on Machine Learning , pages=

    Explaining time series predictions with dynamic masks , author=. International Conference on Machine Learning , pages=. 2021 , organization=

    2021

  78. [78]

    International Conference on Machine Learning , pages =

    Learning Perturbations to Explain Time Series Predictions , author =. International Conference on Machine Learning , pages =. 2023 , volume =

    2023

  79. [79]

    International Conference on Machine Learning , pages =

    Axiomatic Attribution for Deep Networks , author =. International Conference on Machine Learning , pages =. 2017 , month =

    2017

  80. [80]

    2014 , eprint=

    Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps , author=. 2014 , eprint=

    2014

Showing first 80 references.