acceptodds
Under review as a conference paper at ICLR 2027

Temporal Hide&Seek: Instance-wise Time-Feature Selection for Time Series

Abstract

Instance-wise feature selection seeks to identify inputs that are informative for individual predictions. For time series, this requires identifying not only which variables matter, but also when they matter. This creates distinct challenges: the relevance of an observation can depend on its temporal context, the relevant cells can change from one window to the next, and masking individual observations can disrupt the temporal structure the predictor relies on. We introduce Temporal Hide&Seek (T-HiSe), a sequence-aware method for instance-wise selection of time–feature pairs that learns a selector and a predictor jointly and end to end. A Transformer-based Hide network reads the full input window to produce a continuous mask over every time–feature pair, and an LSTM-based Seek network predicts from the masked sequence. Masked observations are replaced with complete feature trajectories drawn from strictly earlier training windows, preserving each feature’s temporal structure while avoiding future information, and an annealed sparsity penalty sharpens the masks as training progresses. On eight synthetic benchmarks with known ground-truth supports, spanning fixed and context-dependent supports, T-HiSe attains the best or tied-best support recovery on seven and averages 97.2%, with the largest gains where the relevant support switches between windows. In two real-world case studies, hourly bike-sharing demand and EEG seizure detection, T-HiSe achieves predictive performance comparable to the reported references while assigning high mask values to a relatively small part of each window. Deleting these highly ranked cells reduces predictive performance substantially more than deleting the same number of random cells.

Primary area:
interpretability and explainable AI
Keywords:
Instance-wise feature selection, multivariate time series, explainable AI, temporal feature selection, differentiable masking, Transformers, LSTMs, Hide&Seek.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_rFSVOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Temporal Hide&Seek (T-HiSe), which extends Hide&Seek to instance-wise selection of time–feature cells in time series. A Transformer Hide network outputs a continuous mask over all cells of the input window. An LSTM Seek network predicts from a mixed window, in which hidden cells are blended with feature trajectories copied from earlier training windows. Both networks are trained jointly with an annealed sparsity penalty.

Critical strengths
  • The synthetic benchmarks provide exact ground-truth supports, including context-dependent supports that switch between windows.
  • The benchmark generation is transparent. The data-generating processes were fixed before model tuning, and code is provided for all experiments.
Critical weaknesses
  • The problem definition and terminology are ambiguous. It is unclear what “this” refers to in the second paragraph of the introduction. “Cell” is used without definition until it appears as a time-feature pair in the contributions. The stated goal mixes explaining an existing predictor with selecting features by jointly training a selector and a predictor. “Relevant” is not defined with respect to either the label or the predictor.
  • The claim that time series are harder than tabular data is weakly supported. Of the three properties, window-dependent supports and cells that cannot be judged in isolation apply equally to instance-wise selection on tabular data. The only time-series-specific difficulty is that replacement values break autocorrelation. No evidence shows that existing methods actually fail.
  • The problem is hard to regard as unsolved. The introduction does not discuss time-series methods that already learn time–feature masks with temporal perturbations, such as DynaMask, ExtrMask, and ContraLSP. The contribution is therefore limited.
  • The link between the challenges and the solution is weak. Any selector that reads the full input already accounts for context, so the Transformer is an architecture choice rather than a new solution. Replacing hidden cells with trajectories from earlier windows creates discontinuities at the boundaries with retained values and breaks cross-feature correlations.
  • In joint training, the predictor adapts to the replaced inputs, which weakens the OOD motivation itself. Meanwhile, the known key problem of joint training, namely the selector encoding the prediction in the selection (Jethani et al., 2021), is handled only by two informal safeguards: Seek does not see the mask, and hidden cells are filled with real data. No test is included that could detect such encoding.
  • There is no ablation. It is impossible to tell which of the Transformer Hide network, trajectory-based donor replacement, or annealing produces the gain over Hide&Seek.
  • The baseline comparison is unfair. DynaMask is run without temporal regularization, WinIT’s learned generator is replaced with marginal sampling, and L2X and INVASE use zero replacement. Hide&Seek uses different epoch and λmax⁡\lambda_{\max} settings, and LIME, TimeSHAP, and WinIT explain only 200 windows. ExtrMask (Enguehard, 2023) is cited but not compared. The choice of λmax⁡=9\lambda_{\max} = 9 is not justified.
  • The synthetic data favor the proposed method. Because the features are independent AR(1) processes, independent per-feature donor replacement is exactly in-distribution. Support positions are also fixed relative to the window end. This advantage is not guaranteed on real data with correlated features.
  • On real data, T-HiSe is compared only against random selection, not against other explainers. On EEG, a logistic model on amplitude alone outperforms T-HiSe, and the correlation between the mask and ∣x∣|x| is only 0.28. The added value of the explanation is unclear.
Questions for authors

See the weaknesses.

1 comment

We thank the reviewer for the careful review. We agree with the main points and have revised the paper accordingly.

  • Terminology: "cell" (time–feature pair) and "relevant" are now defined in the introduction and Section 3.1, and we clarify that all experiments use the jointly trained mode.
  • Time series vs. tabular: we removed the claim that selection is harder for time series.
  • Related work: DynaMask, ExtremalMask and ContraLSP are now discussed and included as baselines.
  • Transformer selector: we agree any selector that reads the full input can use context. We chose the Transformer because it provides positional context for every cell in parallel, which keeps mask generation fast.
  • Design: Section 3.3 now acknowledges that trajectory donors can create boundary discontinuities and do not preserve cross-feature dependence.
  • Encoding the prediction in the mask: we agree this is not tested in the current version and note it as a limitation.
  • Ablations: we added a controlled ablation of trajectory donors vs. zero replacement with everything else fixed. We agree the Transformer and annealing are not yet ablated.
  • Sparsity weight: λ was initially chosen on separate developmental datasets. In the revision we follow the protocol of the ContraLSP paper: all other hyperparameters are fixed, and λ is selected from {0.1, 1} as the value that maximizes validation accuracy, on a separate seed not used for evaluation (Appendix).
  • Baselines: WinIT and FIT use their learned generators, ExtremalMask is added, and every method explains every test sequence.
  • Synthetic data: we now use standard benchmarks from the literature (rare-feature, rare-time, State, Switch-feature). We agree per-feature donors suit independent features.
  • Real data: the EEG study is removed, and on bike sharing T-HiSe is now compared with ContraLSP under the same deletion/retention protocol.

Thank you again for the helpful feedback.

Reviewer_Jbw1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes T-HiSe, a method for instance-wise selection of time-step/feature pairs in multivariate time series. A Transformer ("Hide") reads the full window and outputs a continuous mask over every cell; an LSTM ("Seek") predicts from the masked window; the two are trained jointly under an annealed sparsity penalty on the mean mask. Masked cells are not zeroed: for each feature, the whole column is copied from another training window that ends before the target window, so replaced cells carry a real trajectory and no observation after the prediction time enters the input.

The method is evaluated on eight synthetic datasets whose (time-step, feature) supports are known exactly — two global, six with supports gated on part of the window — against eight baselines spanning post-hoc attribution (Integrated Gradients, LIME, TimeSHAP, WinIT, DynaMask) and learned selection (L2X, INVASE, and the original Hide&Seek applied to flattened windows). Support recovery is measured as top-k recall with k equal to the true support size, averaged over five seeds. T-HiSe attains the best or tied-best mean on seven of the eight datasets, averaging 97.2% against 73.4% for the next-best method. Two case studies — hourly bike-sharing demand and EEG seizure detection — report predictive performance close to reference models, deletion and retention gaps against random cells, and domain-level descriptions of what the masks select.

Critical strengths
  1. The synthetic benchmarks give a controlled setting in which support recovery can be measured rather than argued about. The eight generators specify the relevant (time step, feature) cells exactly, and they cover two-way and three-way switches, a gate over a subsequence, a comparison between adjacent blocks, correlated decoys and a cell twenty-two steps back. Two of the gates are aggregates over several cells rather than single values, so a correct explanation has to select a set. This matters for my assessment because it is the only part of the paper where the target is known, and everything the paper claims about instance-wise behaviour rests on it.

  2. The protocol around temporal leakage is more careful than is usual in this area, and it is what makes the case studies interpretable. The EEG data are split by recording, the bike-sharing split is chronological with a gap of one window length so that no hour falls in two splits, and training donors are restricted to windows ending before the prediction time, with the first-window case and the test-time change of donor pool both stated rather than glossed. In the deletion and retention tests, the same masks and the same donor draws are applied to the mask-based and the random conditions, so the comparison isolates which cells were ranked highly from what was substituted for them. I regard the splitting and donor choices as the most consequential part of this protocol, and they are handled well.

Critical weaknesses
  1. I am not convinced that the comparison isolates the contribution. As submitted, I cannot tell how much of the reported gap on the synthetic benchmarks comes from the temporal selector and how much from the replacement scheme, which in that comparison is applied only to T-HiSe.
  • The baselines are run without the operators that define them. Section 3.3 argues that replacing hidden cells with a constant, or with an independent random value per cell, "breaks the temporal structure of the series and gives Seek inputs unlike real data". Table 5 then runs L2X-adapted and INVASE with unselected cells set to zero, and Hide&Seek with per-cell marginal draws, while only T-HiSe receives trajectory donors. WinIT is the clearest case: Section 2 credits it with "accounting for the dependence between successive observations of the same feature", and Table 5 removes the learned conditional generator that does exactly that ("We use a generative approach to replace the removed observations", WinIT Section 3.2). L2X's jointly learned variational approximator is likewise replaced by a frozen black box. I would like to see at least one baseline given a replacement scheme comparable to T-HiSe's.

  • The closest prior work among the citations is not evaluated. Enguehard (2023) learns a mask over "which features at which times" together with the perturbation itself, and it is absent from Table 5, while L2X and Hide&Seek are run on flattened windows.

  • The added mechanism is never ablated. There is no variant with per-cell marginal replacement, none with a feed-forward selector, and none with a constant replacement. I think this comparison is worth doing here in particular, because the method T-HiSe extends already compares replacement schemes directly (Appendix A.2, marginal against knockoff) and states the condition under which a replacement is leakage-free (Appendix A.3). Neither is mentioned in the submission, and no equivalent is given for the trajectory variant.

  • The leakage argument is inherited rather than tested. That Seek never sees the mask, and that hidden cells are filled with real data, are the arguments of the method this paper extends. Oosterhuis et al. (2024, Corollary 2.5) give a necessary and sufficient condition for a selector without leakage: the probability of a selection must depend on no label values and on no non-selected feature values. T-HiSe's premise is that the relevance of a cell depends on its temporal context. I do not see how both can hold without an explicit trade-off, and the paper does not discuss it.

  • Three reporting choices make the numbers harder to interpret than they need to be. No accuracy is reported for the LSTM that the post-hoc methods explain, although it is described as "well-trained". LIME, TimeSHAP and WinIT are scored on 200 windows while every other method explains the full test set, and the text still states that "All methods are scored on the same test windows". Hide&Seek is run for 500 full-batch epochs, so its number of parameter updates differs from T-HiSe's by roughly two orders of magnitude rather than by a factor of two.

  • The reproducibility statement promises material that is not there. It states that the implementation, the generators for all eight synthetic datasets, the scripts for every baseline and the two case-study pipelines "reproduce the splits and write every reported number", and that these are provided in the anonymized supplementary material. No supplementary material is attached to the submission and the paper gives no link to one, so the claim that every reported number is reproduced by a released package cannot be checked by a reviewer.

  • The sparsity weight has no stated provenance, and the margin is narrower than it reads. syn1–syn7 have no validation split and syn8's validation block is "not used by any method", so it is unclear on what data the single value of 9 was chosen; Appendix A.6 shows that reducing it to 2 moves syn7 from 90.2 to 94.0 and pulls Hide&Seek from 90.6 to 87.3, reversing that comparison. Two of the eight columns are also at ceiling, with four or five methods tied at 100.0, so "+23.8 points" is a distance to an upper bound rather than an effect size.

References cited: Leung et al., "Temporal Dependencies in Feature Importance for Time Series Prediction", ICLR 2023; Chen et al., "Learning to Explain: An Information-Theoretic Perspective on Model Interpretation", ICML 2018; Yoon et al., "INVASE: Instance-wise Variable Selection using Neural Networks", ICLR 2019; Ellinson et al., "Hide&Seek: Learning to Explain in an End-to-End Differentiable Network", ICML 2026; Enguehard, "Learning Perturbations to Explain Time Series Predictions", ICML 2023; Oosterhuis, Lyu & Anand, "Local Feature Selection without Label or Feature Leakage for Interpretable Machine Learning Predictions", ICML 2024; Ribeiro et al., "'Why Should I Trust You?': Explaining the Predictions of Any Classifier", KDD 2016; Bento et al., "TimeSHAP: Explaining Recurrent Models through Sequence Perturbations", KDD 2021.

  1. The case studies measure a weaker property than the paper claims. They show that a coarse, window-averaged ranking is predictive, and that deleting high-mask cells hurts, but not that the selection is instance-specific, and not that the harm comes from removing information rather than from presenting the model with an input it does not recognise.
  • Every piece of real-data evidence is a window-averaged summary: per-feature mean masks and a lag-decay curve for bike sharing, and per-segment means, one within-segment correlation with the input magnitude and one group contrast for EEG. No window-level selection is shown, so what the paper demonstrates on real data is feature- and lag-level structure, that is, a global explanation; the abstract's claim that "the relevant support switches between windows" is tested only on the synthetic gates.

  • The only control in Table 2 is random cells. Model randomisation (Adebayo et al., 2018) and retraining after removal (Hooker et al., 2019) are both absent, so I cannot separate a drop caused by removing information from a drop caused by an input the model does not recognise — the effect the paper itself invokes in Section 3.3 to justify its own replacement scheme.

  • The readings are not monotone, and only one end is discussed. Deleting 7 of 360 cells takes bike-sharing R² from 0.725 to −0.672, below a predictor of no change, while keeping the top 20% of cells gives 0.735, above the full model; the text records the latter as "at least as good as the full model" without comment. An intervention that improves the metric when 80% of the input is removed does not yet measure faithfulness.

  • The paper contains one figure, the architecture diagram, and no mask visualisation: no heat map, no per-window example, no waveform overlay. The claimed patterns can therefore be checked only through the summary statistics reported in the text, not against any individual window, which is the unit the instance-wise claim is about.

  • On bike sharing the features are not identified. Total rentals are casual plus registered inside the log transform, so the reported split of 0.14 on registered rentals and 0.13 on total, with casual rentals not reported at all, is not identified among near-collinear inputs, and the paper does not say so.

Taken with the first weakness, the two halves of the paper also use different notions of a correct explanation: recovering the generative cells on synthetic data, where the gate cell is counted into the support size, versus sufficiency on real data. The abstract's "best support recovery" is the first of these, not the second.

Until the comparison is rerun with each baseline under its own replacement operator, the closest prior work is evaluated, and the trajectory donor is isolated by ablation, the central claim of the paper is not established.

References cited: Adebayo et al., "Sanity Checks for Saliency Maps", NeurIPS 2018; Hooker et al., "A Benchmark for Interpretability Methods in Deep Neural Networks", NeurIPS 2019.

Questions for authors
  1. In Table 1, the post-hoc methods explain an LSTM that Section 4.1 describes as "well-trained", but no accuracy is reported for it. What are its errors on syn1–syn8, and the errors of T-HiSe's own predictor? On syn4 in particular, where no baseline exceeds 58.3%, I cannot tell whether the low scores reflect the attribution procedures or the predictor they explain.

  2. Table 5 replaces WinIT's learned conditional generator with marginal sampling and L2X's variational approximator with a frozen black box. What happens if T-HiSe's trajectory donor is replaced by per-cell marginal draws, or the Transformer by a feed-forward selector, with everything else fixed? Without that comparison I cannot assign the gain to the temporal structure the abstract credits.

  3. Enguehard (2023) learns a mask over feature-time cells together with the perturbation, which appears to be the closest prior formulation to T-HiSe, but it is not among the baselines. Is there a reason for that? This is a clarification request rather than a suggestion for future work, since it bears on how the contribution is positioned against that work.

  4. Could the authors make the reproducibility package available, or say which reported numbers can be reproduced from the paper alone?

References cited: Leung et al., "Temporal Dependencies in Feature Importance for Time Series Prediction", ICLR 2023; Chen et al., "Learning to Explain: An Information-Theoretic Perspective on Model Interpretation", ICML 2018; Enguehard, "Learning Perturbations to Explain Time Series Predictions", ICML 2023.

1 comment

We thank the reviewer for the thorough and constructive review. We agree with many of the concerns and have revised the paper accordingly.

  • Baseline operators: we agree baselines should keep the operators that define them. Each baseline now uses its own perturbation or replacement operator: WinIT and FIT run with their learned conditional generators, and DynaMask uses its temporal regularization on the rare tasks.
  • Closest prior work: we agree ExtremalMask (Enguehard, 2023) should be compared. It is now included as a baseline, together with ContraLSP and DynaMask.
  • Ablations: we added a controlled ablation of trajectory donors vs. zero replacement, with the Transformer/LSTM pair, training budget and selection protocol fixed. We agree that marginal replacement and a feed-forward selector are not yet ablated.
  • Transformer selector: we chose the Transformer because it provides positional context for every cell in parallel, which keeps mask generation fast.
  • Leakage: we agree the leakage argument is inherited rather than tested, and that a context-dependent selector cannot satisfy the condition of Oosterhuis et al. exactly. We note this as a limitation.
  • Reporting: the predictive performance of all black boxes (LSTM(32), GRU(32), GRU(200)) and of T-HiSe's predictor is now reported, the black box explained by each method is listed in the appendix, and every method explains every test sequence.
  • Sparsity weight: λ was initially chosen on separate developmental datasets. In the revision we follow the protocol of the ContraLSP paper: all other hyperparameters are fixed, and λ is selected from {0.1, 1} as the value that maximizes validation accuracy, on a separate seed not used for evaluation (Appendix). We also report AUPRC, AUROC, AUP and AUR in addition to top-k recall, so the comparison no longer depends on a metric with ceiling effects.
  • Case studies: on EEG, our intended claim was that T-HiSe retained only a small fraction of the input while keeping its accuracy. We agree, however, that its accuracy was poor compared with a simple baseline, which made EEG a poor choice of case study, and we have removed it along with the window-averaged per-feature summaries. On bike sharing, T-HiSe is now compared with ContraLSP under the same deletion/retention protocol and donors. We agree this measures sufficiency and necessity rather than instance-specific selection, and that retraining-based controls would strengthen it.

Answers to the questions:

  1. Black-box and T-HiSe predictor performance is now reported for every benchmark.
  2. We added the trajectory vs. zero replacement ablation with everything else fixed; a feed-forward selector ablation remains future work.
  3. ExtremalMask is now evaluated as a baseline.
  4. The revised paper links a notebook that reproduces the experiments.

Thank you again for the detailed feedback.

Reviewer_96RaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Temporal Hide&Seek (T-HiSe), an extension of jointly trained instance-wise feature selection to multivariate time series. A Transformer-based Hide network reads an input window and produces a continuous mask over time–feature cells. An LSTM-based Seek network predicts the target from a mixture of the original observations and replacement values, while an annealed sparsity penalty encourages compact selections.

The main methodological change is a replacement scheme that samples complete feature trajectories from training windows instead of replacing individual cells with constants or independent marginal samples. For chronologically ordered data, donor windows are restricted to those ending before the prediction time.

The evaluation covers eight synthetic benchmarks with known supports, including context-dependent switching tasks, and two real-data case studies on bike-sharing demand and EEG seizure detection. The paper reports an average top-k support-recovery rate of 97.2% on the synthetic benchmarks. The real-data experiments examine predictive performance and the effects of deleting or retaining selected cells relative to random cells.

Critical strengths

1. A well-motivated replacement mechanism for time-series feature selection. The paper addresses a relevant limitation of perturbation-based explanations: replacing observations with constants or independent samples can disrupt the temporal structure used by a sequence model. Sampling complete feature trajectories provides a simple way to retain the temporal structure of the donor signal within each feature. Restricting donors to earlier training windows also gives a clear rule for avoiding future observations in chronological prediction tasks. The approach integrates naturally with continuous-mask optimization and offers a useful direction for adapting selector–predictor methods to time series.

2. Strong synthetic results with informative support-level evaluation. The synthetic benchmarks provide known time–feature supports, allowing the paper to evaluate selection directly in addition to predictive accuracy. The reported gains are particularly substantial on several switching tasks: on syn4, T-HiSe achieves 98.5% support recovery, compared with 58.3% for the strongest reported baseline. I also appreciate the stated decision to fix the data-generating processes before tuning, as well as the analysis of syn7, where the method does not achieve the highest mean recovery. The combination of a simple method and strong results on tasks requiring context-dependent selection is the main reason I support acceptance.

Critical weaknesses

1. Further controlled ablations would clarify the contribution of trajectory replacement. T-HiSe differs from the original Hide&Seek baseline in several respects simultaneously: it uses a Transformer selector instead of an MLP on flattened windows, an LSTM predictor instead of an MLP, trajectory replacement instead of independent marginal replacement, and different sparsity and optimization settings. Appendix A.4 reports 500 full-batch epochs for Hide&Seek, while T-HiSe uses 1000 minibatch epochs. The sparsity weights also differ on syn1–syn7.

These comparisons provide evidence for the effectiveness of the complete framework, but they leave the contribution of the replacement mechanism less clear. A controlled comparison that fixes the Transformer/LSTM pair, optimization budget, and sparsity-selection protocol while varying only the replacement scheme would strengthen the interpretation. Zero replacement and independent marginal replacement would be useful controls alongside complete donor trajectories.

The hyperparameter-selection procedure would also benefit from clarification. The reported synthetic setup does not use a validation set for choosing the sparsity weight, but the paper does not clearly explain how the final values were selected. This matters because Appendix A.6 shows that changing the weight improves syn7 recovery from 90.2% to 94.0%. A description of whether these values were fixed in advance or selected through separate development experiments would help readers assess and reproduce the protocol.

2. The real-data evaluation would benefit from a common independent evaluator. Both real-data case studies compare the selected cells with random cells, without corresponding explanation baselines under the same evaluation protocol. Deletion and retention are also evaluated through the jointly trained Seek network. Because Seek is trained on inputs constructed using Hide’s selections, performance under alternative selections can reflect the predictor’s adaptation to its training masks as well as the information contained in the selected cells.

The reported curves are informative about the observations used by the learned predictor. An evaluator trained independently with random masks and the same donor-replacement mechanism, or a suitably controlled retraining protocol, would provide complementary evidence about the information retained by the selections. This would also help address the selector–predictor evaluation issue discussed by Jethani et al. (2021).

Comparing established explainers under this common protocol would make the real-data conclusions more convincing. The current results support the plausibility and usefulness of the learned selections, while broader comparative claims should remain appropriately qualified.

Overall, I lean toward Weak Accept. The proposed framework is simple, the replacement mechanism is well motivated, and the synthetic results provide encouraging evidence of its effectiveness. The concerns above limit the attribution of individual contributions and the strength of the real-data interpretation, but I find the overall contribution useful enough to support acceptance.

Questions for authors

1.Contribution of trajectory replacement. Could the authors hold the Transformer Hide and LSTM Seek networks fixed and compare donor trajectories, independent marginal samples, and zero replacement? Using matched optimization budgets and a common procedure for selecting sparsity weights would help clarify the contribution of temporally structured replacement.

2. Hyperparameter selection. How were the final sparsity weights for T-HiSe and the baselines selected, given the reported absence of a validation-based selection procedure on the synthetic benchmarks? Were they fixed in advance or chosen using training data or separate development experiments? Clarifying this would help interpret the sensitivity observed on syn7.

3. Independent real-data evaluation. Could the authors compare T-HiSe with at least two explanation baselines on bike sharing and EEG using a common evaluator trained independently with random masks and donor replacement? Reporting variability across training seeds and replacement samples would also help assess the reliability of the deletion and retention curves.

4. Temporal variation beyond fixed templates. The switching benchmarks select among supports at fixed lags relative to the window end. Could the authors discuss how the approach is expected to behave when relevant events occur at variable positions within the input window, and whether the current experiments provide evidence for that setting?

1 comment

We thank the reviewer for the positive and constructive review. We agree with the concerns raised and have revised the paper accordingly.

  • Contribution of trajectory replacement: we agree a controlled comparison is needed. We added an ablation that fixes the Transformer Hide and LSTM Seek networks, the training budget and the sparsity-selection protocol, and varies only the replacement (trajectory donors vs. zero replacement).
  • Hyperparameter selection: λ was initially chosen on separate developmental datasets. In the revision we follow the protocol of the ContraLSP paper: all other hyperparameters are fixed, and λ is selected from {0.1, 1} as the value that maximizes validation accuracy, on a separate seed not used for evaluation (Appendix). Baselines use the hyperparameters reported by their authors for these benchmarks where available.
  • Real-data evaluation: we agree a common protocol with explanation baselines is more convincing. The EEG study is removed, as its accuracy was poor compared with a simple baseline. On bike sharing, T-HiSe is now compared with ContraLSP under the same deletion/retention protocol and donor draws. We agree that each method is still evaluated on its own predictor, and that an independently trained evaluator, as the reviewer suggests, would strengthen this further.
  • Variable positions: the revised synthetic benchmarks are standard benchmarks from the literature, and in several of them the support is not fixed relative to the window end. In rare-time the salient burst starts at a random time, and in State and Switch-feature the relevant feature follows a hidden regime that changes at random times within the sequence.

Thank you again for the helpful feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.