acceptodds
Under review as a conference paper at ICLR 2027

PHaSe: Adaptive Time Series Data Selection via Prototype Hardness

Abstract

Real-world time series data vary in quality, yet conventional training typically grants equal training opportunities to all instances. Although data selection can prioritize useful instances, assessing instances individually does not explicitly account for the overall learning state of their shared temporal patterns. In this work, we propose Prototype HArdness-guided SElection (PHaSe), an adaptive time series data selection framework that aligns training allocation with the model's evolving learning needs across temporal patterns. PHaSe organizes instances into representative pattern groups and estimates their relative learning needs, termed hardness, from recent loss trajectories, providing a pattern-level basis for identifying instances that require further learning. It then transfers pattern-level hardness to instance priorities, modulates these priorities by assignment concentration, and converts them into sampling probabilities. To adapt pattern organization to changing training priorities, we refit prototypes on selected instances and combine the fitted centers with historical centers through normalized momentum updates. Compared with the evaluated baselines in individual settings, PHaSe achieves up to 16.4% lower long-term MSE, 4.8% lower short-term sMAPE, and 74.2% higher classification accuracy, while using only half of the training instances per epoch.

Primary area:
learning on time series and dynamical systems
Keywords:
time series analysis, data selection, prototype hardness

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_NSJEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PHaSe, an adaptive time series data selection method that groups instances into temporal pattern prototypes and estimates their evolving hardness from recent loss trajectories. PHaSe transfers prototype hardness to instance sampling probabilities and dynamically updates the prototypes during training. Experiments across forecasting and classification datasets, multiple backbones, and time series foundation models show consistent improvements over full data training and existing selection methods, supported by ablation, pruning, and runtime analyses.

Critical strengths
  1. Useful pattern level formulation. PHaSe aggregates evolving loss trajectories across related temporal patterns rather than scoring each instance independently. This provides a well motivated and practically simple way to stabilize adaptive selection while accounting for shared structure among time series instances.

  2. Broad and consistent empirical support. PHaSe is evaluated across forecasting and classification tasks, three conventional backbones, and three time series foundation models. Its consistent improvements, together with component ablations, data pruning experiments, five seed evaluation, and runtime measurements, provide substantial evidence that the proposed selection mechanism is effective beyond a single dataset or architecture.

Critical weaknesses
  1. Limited validation under persistent noise. PHaSe prioritizes patterns with sustained high losses, which may also correspond to mislabeled or irreducibly noisy instances. The paper acknowledges this failure mode but does not evaluate it under controlled label noise or temporal corruption, leaving its robustness in the motivating real world setting unclear.

  2. The source of the prototype level gains is not fully isolated. The ablations show that prototype aggregation provides most of the improvement, while loss trend estimation, assignment gating, and dynamic maintenance contribute more modestly. A simple baseline using the same prototypes with uniform or current loss based allocation would help determine whether the full adaptive hardness mechanism is necessary.

Questions for authors
  1. Could the authors report PHaSe’s performance under controlled label noise or temporal corruption?

  2. Could the authors compare PHaSe with a simpler selector that uses the same prototypes but applies uniform allocation or only the current mean prototype loss?

2 comments
AuthorsOpenReview ↗

Response to NSJE: persistent noise and prototype allocation

Thank you for the constructive questions about persistent noise and the source of prototype-level gains.

W1 / Q1. Controlled corruption and persistent high losses.

We agree that persistent high loss can arise from noise or unlearnable examples, so clean-data improvements alone do not establish robustness. We evaluate our method's ability to handle noise on Weather, M4-Monthly, and EthanolConcentration using DLinear, TimesNet, and iTransformer.

We add noise to the training inputs. We first normalize them, add zero-mean, unit-variance Gaussian noise in the normalized space, and transform them back to the original scale before subsequent processing. Corrupted inputs are fixed across epochs. The percentage is the fraction of corrupted training instances. Validation and test data remain clean.

The endpoints are shown below. Appendix E.3, Table 14 contains the complete results at 0%, 1%, 5%, 10%, 15%, and 20% corruption.

BackboneWeather MSE, 0% → 20%Monthly sMAPE, 0% → 20%EC accuracy (%), 0% → 20%
DLinear0.182 → 0.19413.85 → 14.2528.62 → 26.20
TimesNet0.166 → 0.18013.00 → 13.5128.14 → 26.23
iTransformer0.169 → 0.18114.13 → 14.3727.74 → 25.97

Under the tested corruption conditions, our model's performance degradation remains limited as noise increases. One possible explanation is that assignment concentration in Eq. (8) reduces the priority of noisy inputs when their memberships are spread across patterns, making them less likely to be sampled.

W2 / Q2. Uniform and current-loss allocation with the same prototypes.

We evaluate both requested controls. Uniform prototype allocation retains prototype construction and dynamic maintenance, assigns equal sampling mass to prototypes, and distributes it to instances through normalized memberships. It removes loss-based priorities and the assignment gate. Current prototype loss is equivalent to the "w/o loss trend estimation" variant. It replaces the historical trend signal with the latest observation loss and aggregates it through the same soft memberships.

The complete main-text Table 4 is reproduced below. Current prototype loss corresponds to "w/o loss trend estimation", and fixed clustering corresponds to "w/o prototype updates". MSE/MAE are long-term forecasting metrics, sMAPE is short-term forecasting, and accuracy is classification. Entries are mean±standard deviation over five seeds after equal-weight averaging across datasets and backbones within each task.

VariantMSE ↓MAE ↓sMAPE ↓Acc. (%) ↑
Random0.3087±0.00120.3240±0.000710.277±0.02353.35±0.63
Random grouping0.3112±0.01450.3249±0.009810.253±0.08053.37±1.01
Uniform prototype allocation0.3010±0.00030.3129±0.000410.240±0.04053.89±0.41
Instance-level loss smoothing0.3093±0.00090.3243±0.000610.276±0.04453.59±0.53
w/o assignment gating0.2980±0.00060.3082±0.000410.190±0.04754.55±0.46
w/o loss trend estimation0.3004±0.00040.3116±0.000310.220±0.03254.19±0.35
w/o momentum smoothing0.3012±0.00080.3130±0.000610.214±0.05854.28±0.42
w/o prototype updates0.3033±0.00040.3167±0.000310.241±0.03553.92±0.37
w/o prototype aggregation0.3100±0.00110.3253±0.000810.283±0.06453.20±0.58
PHaSe0.2970±0.00050.3076±0.000310.186±0.05154.71±0.43

PHaSe outperforms both requested controls on all three tasks. Compared with uniform allocation, it reduces MSE by 0.0040 and sMAPE by 0.054, and improves accuracy by 0.82 percentage points. It also improves all three task metrics over current-loss allocation, supporting the combination of prototype maintenance, loss-driven priorities, gating, and loss history.

The extended ablation also compares random grouping and instance-level loss smoothing. Random grouping replaces k-means while retaining the other steps. Instance-level loss smoothing takes the arithmetic mean of each instance's latest W losses without prototype aggregation. PHaSe also outperforms these simpler alternatives. Among the five component removals, removing prototype aggregation causes the largest degradation, supporting pattern aggregation for data selection.

The full noise protocol and results are in Appendix E.3, Table 14. The unified ablation is Table 4.

We hope these clarifications and additional results address your remaining concerns. If so, we would be grateful if you would consider a more favorable overall recommendation.

Thanks the author for the comments. My main concerns have been addressed and I will maintain my score.

Reviewer_i2HWOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates adaptive data selection for time series training, arguing that existing instance-level selection methods do not sufficiently exploit the shared temporal patterns and evolving learning states among related instances. To address this limitation, the authors propose PHaSe, which constructs temporal pattern prototypes, estimates their hardness from historical loss trajectories, and transfers pattern-level learning states into instance sampling priorities through soft assignments and dynamic prototype updates. Extensive experiments on long-term forecasting, short-term forecasting, classification, and time series foundation model finetuning demonstrate improvements in downstream performance, with reduced training time in several settings.

Critical strengths
  1. Practically relevant and well-motivated problem. The paper addresses how training resources should be dynamically allocated according to a model's evolving learning state. The observation that instance-level selection may overlook shared information among samples with similar temporal patterns provides a reasonable motivation for pattern-level selection.

  2. Coherent and model-agnostic framework. PHaSe combines prototype aggregation, loss trajectory estimation, soft assignments, and dynamic prototype updates to derive instance sampling priorities from pattern-level learning states. The framework can be applied to different forecasting and classification architectures without modifying their backbones.

  3. Extensive empirical evaluation with useful ablations. The experiments cover long-term forecasting, short-term forecasting, classification, and foundation model finetuning across multiple architectures. Comparisons with Full, Random, TimeInf, and AdaRho demonstrate competitive performance, while ablations of loss trends, prototype aggregation, and dynamic updates provide evidence for the effectiveness of the overall framework.

Critical weaknesses
  1. Limited differentiation from existing adaptive and group-based data selection methods. The central idea of PHaSe is to aggregate historical instance losses through temporal pattern prototypes and use the resulting hardness scores for adaptive sampling. However, clustering-based curricula, group-aware data selection, hard example mining, and importance sampling are well-established research directions. The paper primarily compares against instance-level approaches such as TimeInf and AdaRho, without sufficiently evaluating more direct group- or cluster-based selection alternatives. Consequently, the unique contribution of prototype-level hardness estimation over existing grouped selection mechanisms remains unclear.

  2. Insufficient evidence that prototype hardness reflects the actual benefit of additional training. The proposed method assumes that patterns with higher trend-adjusted losses require more training opportunities. However, persistently high losses may also result from noisy, anomalous, or intrinsically unlearnable instances, for which additional gradient updates may provide little or even negative generalization benefit. The reported downstream improvements and data pruning experiments only indirectly support the utility of the selected instances and do not establish that prototype hardness predicts the marginal benefit of further optimization. More direct evidence is needed to validate this central assumption.

  3. Lack of direct validation that prototypes capture shared learning states. PHaSe groups instances using backbone-independent temporal descriptors and K-means clustering, but similarity in input characteristics does not necessarily imply similarity in learning difficulty or loss trajectories. Although the ablation results support the overall prototype-guided mechanism, they do not directly demonstrate that instances assigned to the same prototype exhibit more coherent learning dynamics. Analyses of within- and between-prototype loss trajectory similarity, together with comparisons against simple instance-level loss smoothing, fixed clustering, and random grouping, would strengthen the justification for pattern-level aggregation.

  4. The claimed efficiency benefits require a more careful analysis. Although PHaSe uses approximately half of the training instances for gradient updates in each epoch, it still performs a full-data forward pass after every epoch to refresh observation losses and additionally maintains dynamic prototypes. These operations introduce non-negligible computational and memory overhead. The reported wall-clock results show that efficiency gains depend substantially on the backbone and dataset, with some settings providing only modest savings or even slightly higher training time than full-data training. Evaluating accuracy–efficiency trade-offs across different selection ratios and comparing against Full/Random training under matched wall-clock or gradient-update budgets would provide stronger evidence of practical efficiency.

  5. The necessity and robustness of several design components remain insufficiently established. While the complete method achieves the best macro-average ablation results, the improvements from assignment gating and loss trend estimation are relatively small, raising questions about their practical significance and stability. Furthermore, PHaSe introduces several hyperparameters, including prototype count, assignment temperature, hardness temperature, loss history length, and momentum. A more systematic sensitivity analysis would help establish whether the observed improvements are robust across datasets and training configurations.

Questions for authors

Q1. Does prototype hardness predict marginal training utility? Can the authors directly examine the relationship between prototype hardness and the improvement obtained from additional gradient updates? In particular, how does PHaSe behave when high-loss prototypes contain noisy or unlearnable instances?

Q2. Why is prototype-level aggregation necessary? Can the authors compare PHaSe with simpler group-based sampling strategies and analyze whether instances within the same prototype exhibit more similar loss trajectories than instances from different prototypes? Such evidence would help establish the value of organizing learning states at the pattern level.

Q3. What is the actual accuracy–efficiency trade-off? Since PHaSe performs full-data loss evaluation after every epoch, can the authors provide comparisons under matched wall-clock time and gradient-update budgets, along with sensitivity analyses across different sampling ratios and prototype update frequencies?

6 comments
AuthorsOpenReview ↗

Response to i2HW (1/5): structured selection and runtime

Thank you for the detailed review and constructive suggestions, which help us improve the paper.

W1 / Q2. Comparison with structured selection methods.

PHaSe pools loss histories into an evolving pattern-level state and uses it for training allocation, while updating prototypes from the selected training instances. We agree that instance-level baselines alone do not isolate the value of grouping, so we add D² Pruning as a structured selection baseline.

D² combines difficulty and diversity through graph message passing. For time series, we construct its graph from backbone representations and initialize difficulty using the task-specific instance observation loss. We retain forward difficulty propagation and reverse neighbor downweighting, select 50% of instances after warm-up, and train on that fixed subset. We use released-code defaults, with unspecified shared training settings taken from the corresponding PHaSe configuration.

The following extracts compare D² and PHaSe directly. Full comparisons with all methods are in main-text Tables 1 and 2. DL/TN/iT denote DLinear/TimesNet/iTransformer. The metric is MSE for long-term forecasting, sMAPE for M4, and accuracy (%) for classification.

DatasetDL D²DL PHaSeTN D²TN PHaSeiT D²iT PHaSe
ETTh10.3990.3950.4110.4080.3870.380
ETTh20.3060.3010.3150.3100.2980.296
ETTm10.3360.3310.3310.3280.3320.326
ETTm20.1830.1820.1710.1700.1770.176
Weather0.1830.1820.1680.1660.1710.169
Electricity0.1950.1930.1620.1600.1480.147
Traffic0.6450.6410.5910.5860.3970.390
M4-Yearly14.1114.0213.6013.5413.8313.77
M4-Monthly14.0213.8513.2313.0014.2514.13
M4-Daily3.012.963.083.053.393.35
JapaneseVowels96.6996.9695.3195.5097.1297.40
Handwriting15.3115.5721.4621.7620.2320.43
Heartbeat73.1273.9375.5275.6574.6274.85
EthanolConcentration27.9728.6227.9928.1427.5827.74

Training time in seconds (mean ± standard deviation):

DatasetBackboneD² (s)PHaSe (s)
ETTh1DLinear19.4 ± 1.817.5 ± 0.2
ETTh1TimesNet91.6 ± 1.582.4 ± 0.3
ETTh1iTransformer32.3 ± 0.527.9 ± 0.1
WeatherDLinear43.8 ± 2.935.7 ± 0.2
WeatherTimesNet231.6 ± 54.2212.9 ± 29.9
WeatheriTransformer78.4 ± 8.865.9 ± 5.0

PHaSe has better mean performance in all 42 combinations and lower mean training time in all six runtime settings. These comparisons support its effectiveness against a structured selector. The subsequent controls further examine grouping and the allocation signal.

AuthorsOpenReview ↗

Response to i2HW (2/5): training utility and persistent noise

W2 / Q1. Does hardness correspond to the benefit of additional updates?

To address this question, we compare checkpoint branches on ETTm1, M4-Monthly, and Heartbeat. For each dataset, we randomly select three checkpoints for each of the three backbones. From the same checkpoint, each branch takes 20 additional parameter updates using low-, medium-, or high-hardness groups, or the complete PHaSe policy. Each dataset/policy entry pools the nine checkpoint-level improvements and reports their mean and standard deviation.

Dataset / improvementLow hardnessMedium hardnessHigh hardnessPHaSe
ETTm1 / MSE reduction0.0343 ± 0.00920.0669 ± 0.04260.0803 ± 0.06120.0860 ± 0.0630
M4-Monthly / sMAPE reduction0.0068 ± 0.00740.0118 ± 0.01060.0144 ± 0.00890.0243 ± 0.0131
Heartbeat / accuracy gain (pp)0.768 ± 0.5250.701 ± 0.1491.374 ± 0.2071.717 ± 0.790

These are absolute metric differences. Positive values indicate improvement. Forecasting uses initial minus final error, and classification uses final minus initial accuracy. High-hardness groups have the largest group-wise mean on all three datasets. The complete policy has the largest overall mean, supporting hardness-guided training allocation. These results also appear in main-text Table 5.

We additionally evaluate our method's ability to handle noise on Weather, M4-Monthly, and EthanolConcentration. Only training inputs are corrupted. We normalize each selected input, add zero-mean, unit-variance Gaussian noise, and apply inverse normalization before subsequent processing. Corrupted inputs are fixed across epochs. The percentage is the fraction of corrupted training instances. Validation and test data remain clean.

The endpoints are shown below. Appendix E.3, Table 14 contains all six corruption levels: 0%, 1%, 5%, 10%, 15%, and 20%.

BackboneWeather MSE, 0% → 20%Monthly sMAPE, 0% → 20%EC accuracy (%), 0% → 20%
DLinear0.182 → 0.19413.85 → 14.2528.62 → 26.20
TimesNet0.166 → 0.18013.00 → 13.5128.14 → 26.23
iTransformer0.169 → 0.18114.13 → 14.3727.74 → 25.97

From 0% to 20% corruption, Weather MSE rises by 6.6%–8.4%, Monthly sMAPE by 1.7%–3.9%, and EC accuracy falls by 1.77–2.42 percentage points across backbones. In these settings, the observed noise-induced performance degradation of our method remains limited. One possible explanation is that assignment concentration in Eq. (8) reduces the priority of noisy inputs when their memberships are spread across patterns, making them less likely to be sampled.

AuthorsOpenReview ↗

Response to i2HW (3/5): shared loss dynamics

W3 / Q2. Shared loss dynamics.

For each grouping, we calculate Pearson correlations between instance loss trajectories using the same history window W as the trend estimator. For this diagnostic, each instance is assigned to its largest-membership prototype, ci=arg⁡max⁡kqikc_i = \arg\max_k q_{ik}. Pairs with the same label are within-group pairs, and the rest are between-group pairs. Training and selection retain soft memberships.

Main-text Table 6 reports within-group and between-group correlations separately for DLinear, TimesNet, and iTransformer. Within each backbone, long-term forecasting, short-term forecasting, and classification average 7, 3, and 4 datasets equally, respectively. The complete results from Appendix E.2, Table 13 are reproduced below. DL/TN/iT denote DLinear/TimesNet/iTransformer. JV/HW/HB/EC denote JapaneseVowels/Handwriting/Heartbeat/EthanolConcentration. Yearly/Monthly/Daily are the M4 subsets.

DatasetDL withinDL betweenTN withinTN betweeniT withiniT between
ETTh10.6880.3460.8550.3260.7180.448
ETTh20.6830.4080.7270.4920.8040.351
ETTm10.8440.3990.7360.4460.6950.451
ETTm20.8020.4040.8070.4540.8280.428
Weather0.8540.3410.8090.4080.7030.430
Elec.0.7030.4220.7930.4950.7570.339
Traffic0.7490.3490.7230.4230.6880.458
Yearly0.7460.3180.7770.4600.7090.311
Monthly0.7850.4570.6900.3680.8450.429
Daily0.8110.4820.7820.3830.7760.471
JV0.7360.4430.7400.3180.6890.385
HW0.8140.3130.7500.3370.6840.392
HB0.7070.3520.8380.3140.7840.363
EC0.7600.4420.6870.4950.7620.482

Within-group correlations exceed between-group correlations in all 42 dataset–backbone settings, showing coherent loss trajectories among grouped instances.

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

The paper proposes PHaSe, a dynamic data selection method guided by prototype hardness. It constructs handcrafted descriptors of time series inputs and groups them using k-means. Following model warm-up, recent loss histories are aggregated through prototype memberships assignment, and to estimate pattern-level hardness. These hardness scores guide probabilistic selection of training instances for the next epoch, with sampling-based loss reweighting during optimization. Prototypes are refitted on the selected instances and blended with historical centers through momentum updates.

The paper provides implementation and benchmarking details, evaluates several time series tasks and model architectures, and includes ablation studies.

A relevant related work is PEAKS (Gurbuz, Zheng, and Dovrolis, ICML 2025), which derives a theoretically motivated data selection criterion combining current prediction error with similarity to class prototypes. Although the task and selection setting differ, discussing this connection would help contextualize PHaSe’s use of prototype structure and model-dependent learning needs.

Critical strengths
  1. The results are clearly presented across forecasting and classification tasks and multiple backbone architectures.

  2. PHaSe achieves good improvements in long-term forecasting over the evaluated baselines. The ablations also provide useful evidence to judge the relevance of each component of the model.

Critical weaknesses
  1. The method introduces many symbols and several interacting practical steps. A simpler overview distinguishing the core idea from supporting design choices would improve readability.

  2. The conceptual novelty needs to be clarified relative to existing model-dependent data selection methods.

Questions for authors

What motivates the particular handcrafted descriptor and the choice of k-means? Would the authors provide a comparison with an alternative representation or grouping method to assess how strongly the results depend on these choices?

3 comments
AuthorsOpenReview ↗

Response to RYiT (1/2): mechanism and structured selection

Thank you for recognizing the empirical coverage and ablations, and for identifying where the presentation and positioning can be clearer.

W1. Core mechanism and supporting design choices.

PHaSe's central mechanism is to aggregate instance loss trajectories into a pattern-level learning state and use it to allocate subsequent training opportunities. Descriptors and clustering organize the patterns, assignment concentration regulates the transfer from prototype hardness to instance priorities, and dynamic updates and momentum maintain the prototypes. Section 3.1 now introduces the intuition before the mathematical details: a model's mastery varies across temporal patterns and evolves during training, so PHaSe constructs prototypes, estimates their hardness from loss trends, and uses this state to guide subsequent selection.

W2. Positioning relative to model-dependent selection, including PEAKS.

We agree that using model-dependent information for data selection is an established direction. Our contribution is how this information is organized. PHaSe shares loss-trajectory information through soft prototype memberships, estimates a pattern-level state, and transfers it back to instance priorities. The prototypes are updated from the selected training instances. Clustering or model-dependent selection alone is not our claimed novelty.

PEAKS is a relevant connection. It combines prediction error with class-related feature alignment in an incremental selection setting. PHaSe instead uses unsupervised temporal-pattern groups to pool loss histories and repeatedly allocate training. We compare PEAKS on four time-series classification datasets and all three backbones. The adaptation retains its scoring, class-count correction, dynamic acceptance threshold, and incremental retention mechanism, with 50% of the training instances retained. We use released-code defaults, and unspecified training settings shared with PHaSe follow the corresponding PHaSe configuration.

PHaSe has higher mean accuracy in all 12 combinations, with improvements of 0.06–1.93 percentage points.

The public implementation and scoring rule of PEAKS use class labels, class-related logits, and class counts, making it naturally adaptable to time-series classification. Applying it to value forecasting would require an additional regression scoring rule, which is beyond the scope of direct reproduction. We therefore report classification comparisons in this response and do not include PEAKS in the main-text table covering both forecasting and classification.

The PEAKS comparison is shown below, reporting mean accuracy over five random seeds.

BackboneDatasetPEAKS accuracy (%)PHaSe accuracy (%)
DLinearJapaneseVowels96.9096.96
DLinearHandwriting15.4315.57
DLinearHeartbeat73.8473.93
DLinearEthanolConcentration27.5528.62
TimesNetJapaneseVowels95.3295.50
TimesNetHandwriting20.4721.76
TimesNetHeartbeat75.5175.65
TimesNetEthanolConcentration26.8928.14
iTransformerJapaneseVowels96.6997.40
iTransformerHandwriting19.3320.43
iTransformerHeartbeat73.8074.85
iTransformerEthanolConcentration25.8127.74
AuthorsOpenReview ↗

Response to RYiT (2/2): representations and grouping distances

Q1. Why the current descriptors and k-means, and whether alternatives work.

The handcrafted descriptors provide backbone-independent summaries of temporal profiles, local changes, and magnitude statistics. K-means gives a direct grouping mechanism compatible with the center-based memberships and updates in our implementation. They are implementation choices rather than prerequisites for the idea of pattern-level state aggregation.

We evaluate three representations (handcrafted, MLP, and Transformer) and four grouping methods (k-means, k-Shape, shapelet-based grouping, and DTW k-means). Both learned alternatives take time-series patches as input and are trained separately for each dataset to reconstruct the input itself, then frozen for representation extraction and grouping. The MLP has two layers with dimensions d → d/2 → d. The Transformer has two layers. Thus, these representations are learned through reconstruction rather than taken from untrained random mappings. Here, shapelet-based grouping replaces Euclidean distance with a shapelet-based distance in the grouping procedure. It does not refer to a separately named clustering algorithm.

Across the tested combinations, ETTh1 MSE ranges from 0.393 to 0.399, Weather MSE from 0.171 to 0.174, Yearly sMAPE from 13.77 to 13.79, JapaneseVowels accuracy from 96.59% to 96.63%, and Handwriting accuracy from 19.24% to 19.25%. These consistently small changes across tasks show that PHaSe is insensitive to the tested descriptor and clustering choices. Its effectiveness is therefore not tied to the original handcrafted representation or k-means implementation.

As the complete forecasting and classification results are lengthy, please see Tables 11 and 12 in the appendix. The protocol is in Appendix E.1.

We hope these clarifications and additional results address your remaining concerns. If so, we would be grateful if you would consider a more favorable overall recommendation.

Thank you for the responses and additional experiments. I appreciate the comparison with PEAKS, the explanation of the core mechanism, and the evaluation of alternative representations and grouping methods. These additions address my questions. I continue to consider 3/4 (weak acceptance) a fair rating and remain supportive of acceptance.

Best, Reviewer RYiT

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response: additional evidence and manuscript revisions

We thank the reviewers for their careful reading and constructive feedback. The revised main text is 10 pages and incorporates the additional main comparisons. Repeated training experiments use five random seeds.

  1. Structured selection and efficiency. Table 1 includes D² Pruning across all 42 dataset–backbone combinations, and Table 2 reports its runtime. Table 3 compares all six methods under separate post-warm-up wall-clock and gradient-update limits. PHaSe has the best task-level mean under both constraints.

  2. Components and training utility. Table 4 combines component ablations with random grouping, instance-level loss smoothing, and uniform prototype allocation. The complete framework performs best on all four metrics. Table 5 reports gains from 20 additional updates, pooling three checkpoints from each of three backbones per dataset. High-hardness groups have the largest group-wise mean, and the full policy has the largest overall mean.

  3. Alternatives and sensitivity. Appendix E.1 shows stable performance across the tested descriptors and clustering methods. Main-text Table 6 summarizes loss-trajectory correlations by task and backbone, with complete results in Appendix E.2. Appendix E.3 details Gaussian input corruption that is fixed across epochs. Appendix E.4 groups sampling/update-frequency and internal-hyperparameter sensitivity.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.