Pith. sign in

REVIEW 4 major objections 6 minor 36 references

Bridging Neural Networks and Dynamic Time Warping for Adaptive Time Series Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims a length-shortening algorithm lets the dynamic time warping recurrence be rewritten as a recurrent neural network, yielding a classifier that is interpretable, works in cold-start settings, and trains on labels.

desk verdict Promising heuristic for interpretable TSC, but the 'equivalent RNN' claim is unsupported and Eq. 5 is misindexed; should be reframed and heavily revised. read the letter →

arxiv 2507.09826 v1 pith:CE4YYCKN submitted 2025-07-13 cs.LG cs.AI

Jintao Qu , Zichong Wang , Chenhao Wu , Wenbin Zhang This is my paper · ORCID
classification cs.LGcs.AI
keywords timeseriesclassificationdynamicwarpingrecurrentneuralnetworksinterpretabilitylow-resourcelearningprototypeDTW-RNNequivalence
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that dynamic time warping (DTW), normally a non-parametric distance for time series classification, can be converted into a trainable recurrent neural network without losing its alignment-based, interpretable decision logic. The conversion works through a length-shortening algorithm that merges nearby coordinates of each class prototype; after shortening, the horizontal one-to-many move in the DTW recurrence can be dropped, leaving a two-way recurrence that is implemented as a recurrent layer. The resulting model is initialized from prototype instances, behaves like nearest-neighbor DTW in cold-start settings, and can be trained end-to-end when labeled data arrives. If correct, this gives a single classifier that is transparent and strong in low-resource settings and competitive with deep learning baselines when data is plentiful.

What carries the argument

The central object is the simplified DTW recurrence that the paper obtains after dynamic length shortening. Length shortening iteratively merges the closest successive coordinates of a prototype, and the paper argues that this mostly eliminates the horizontal one-to-many warping move, leaving only the downward and diagonal moves. That two-way recurrence is then identified with a recurrent neural network: the pairwise distance term acts as the input transformation, max-pooling over the negated padded hidden state computes the min, and the final hidden state combined with a one-hot vector extracts the DTW cost per prototype. The length-shortening step is what carries the argument: it shrinks the parameter space, removes the need to keep a full cost matrix, and is the justification for the claimed equivalence.

What would settle it

Take a dataset whose optimal warping paths contain long horizontal runs, such as sequences with repeated constant plateaus, shorten the prototypes with the paper's algorithm, and compare exact DTW distance with the final cost $h_N[L]$ returned by the simplified recurrence; a systematic relative gap between the two would show the simplification is not equivalent to DTW.

Watch

Extended reading notes

Core claim

The central claim is that DTW and recurrent networks are formally bridgeable for classification: after shortening prototypes, the DTW state update $h_t[j] = \Delta_t[j] + \min(h_{t-1}[j], h_{t-1}[j-1])$ becomes a recurrent cell over prototype-alignment costs. The model is parameterized by a prototype tensor, an input transform, hidden states, and a one-hot readout; distances are computed in batch, the min operation is implemented with max-pooling over the negated padded hidden state, and class scores aggregate per-class prototype alignments with a soft logical OR. Because prototypes are class instances, predictions can be traced back to reference sequences, and because the whole pipeline is differentiable, the parameters can be trained with labels. The paper reports wins over several baselines on distance-oriented benchmark datasets and best or near-best accuracy on six datasets under 1% to 100% training subsampling.

Load-bearing premise

The load-bearing premise is that merging the closest successive coordinates of a prototype leaves DTW distances nearly unchanged, so a simplified recurrence with only two of DTW's three alignment moves is faithful enough for classification; the paper supports this only with a toy illustration and gives no error bound or ablation.

Editorial extensions

If this is right

  • In cold-start settings, the model can be initialized from class prototypes and used immediately, matching nearest-neighbor DTW behavior without gradient training.
  • With more labels, the same architecture is trainable end-to-end, so a deployment can start instance-based and improve as data accumulates instead of being replaced by a new model.
  • The recurrent formulation replaces the full cost-matrix dynamic programming of DTW with batched tensor operations, so the alignment computation becomes parallelizable.
  • Interpretability is inherited: each prediction can be traced to aligned prototypes, and learned prototypes can be inspected, which the paper argues suits high-stakes domains such as medical diagnosis.
  • Because the simplified recurrence drops one of DTW's three moves, datasets whose optimal alignments rely heavily on one-to-many matches may see classification quality diverge from true DTW.

Reading between the lines

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

  • A test the paper does not report is to measure, dataset by dataset, the relative error between exact DTW distance and the value returned by the simplified recurrence after shortening; small error would empirically support the claimed equivalence, whereas large error would imply the wins come from learned parameters rather than DTW fidelity.
  • The layer resembles a differentiable alignment module, so it could be reused outside classification, for example in clustering, retrieval, or contrastive metric learning on sequences.
  • Because the prototype coordinates are fixed while only the input transform is trained, one could test whether making the prototype coordinates themselves learnable improves rich-resource accuracy further, at some cost to the instance-based interpretation.
  • If the equivalence is tight, DTW's alignment bias could serve as an inductive bias for deep sequence models generally, giving a rare continuous path from an interpretable non-parametric method to a deep network.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a trainable recurrent neural network for time series classification that is claimed to be an equivalent reformulation of dynamic time warping (DTW). The core idea is a dynamic length-shortening algorithm that compresses prototype sequences by merging similar successive coordinates, after which the DTW recurrence is approximated by a two-move recurrence (downward and diagonal only). This recurrence is then recast as an RNN with trainable prototype parameters, followed by a soft aggregation layer for classification. The authors evaluate the model on all 85 UCR datasets and on six datasets under subsampled low-resource regimes, reporting strong performance relative to NN-DTW, LSTM, ROCKET, InceptionTime, BOSS, and the Elastic Ensemble.

Significance. If the claimed equivalence were established, the model would be a meaningful bridge between interpretable instance-based methods and trainable neural approaches, with potential value in cold-start time series classification. The paper's broad empirical scope across the UCR archive and its explicit focus on low-resource settings are useful strengths, and the authors provide an anonymous code repository. However, the central formal claim is not currently supported: the abstract and conclusions describe an 'equivalent' RNN, while the body introduces Equation (3) only as an approximation, and the batch recurrence in Equation (5) contains an indexing error that makes the printed network inconsistent with Equation (4). The length-shortening algorithm is also underspecified, and the low-resource performance claims lack statistical support. These issues are load-bearing for the paper's main contribution, so the manuscript needs substantive revision before it can be accepted.

major comments (4)
  1. [Abstract, Section 3.1, Eq. (3)] The abstract and conclusions call the resulting network an 'equivalent' reformulation of DTW, but Section 3.1 introduces Equation (3) only as an approximation after stating that the horizontal move 'tends' to be eliminated. No formal condition, error bound, or empirical comparison between Equation (2) and Equation (3) is provided; the only support is the toy illustration in Fig. 1. This is a load-bearing gap because the claimed DTW-RNN equivalence depends entirely on the validity of dropping the horizontal move. The authors should either provide a precise characterization of when the approximation holds, add a quantitative ablation comparing exact DTW distances with the shortened two-move recurrence on real datasets, or revise all equivalence claims to describe an approximate/heuristic alignment model.
  2. [Section 3.2, Eq. (5) and Eq. (7)] Equation (5) as printed computes min(h_{t-1}[i,j], h_{t-1}[i-1,j]), mixing the prototype index i with the prototype-time index j. The single-prototype recurrence in Equation (4) correctly uses h_{t-1}[j-1], so the batched extension should read min(h_{t-1}[i,j], h_{t-1}[i,j-1]). Equation (7) repeats the same erroneous indexing and does not specify whether the max-pooling is applied along the prototype dimension or the time dimension. As written, the network does not implement the recurrence described in Section 3.1. This must be corrected and the implementation dimension clarified, otherwise the formal claim is not reproducible.
  3. [Section 3.1, Section 4.1] The dynamic length-shortening algorithm is underspecified: 'merge the closest successive coordinates' does not state the distance metric used to choose the pair, the tie-breaking rule, or whether the replacement is the average, the midpoint, or one of the original coordinates. The shortening ratio is selected per dataset by grid search over {0.3, 0.5, 0.8, 0.9}, so the algorithm cannot be replicated or evaluated from the text alone. Since the entire method rests on shortening preserving DTW structure, the paper should provide a precise algorithm, a sensitivity analysis over the shortening ratio, and a direct measurement of how much DTW distance changes after shortening on the benchmark datasets.
  4. [Section 4.3, Abstract] The abstract claims the model 'significantly outperforms previous approaches in low-resource settings,' but this is not supported by the evidence in Table 1. At the 1% sampling rate, the model loses to NN-DTW on ECG5000 (0.840 vs. 0.842) and on Strawberry (0.816 vs. 0.827), and no standard deviations, repeated runs, or significance tests are reported anywhere in the table, despite the text stating that complete results with standard deviations are available online. The authors should report variance across seeds, perform significance testing, and temper the abstract's claim accordingly.
minor comments (6)
  1. [References] References [8] and [9] are the same paper (InceptionTime); this duplication should be removed.
  2. [Section 3.2, Eq. (4)] In Equation (4), h_t is said to be in R^{N+1}, but the recurrence is over prototype length L; this should likely be R^{L+1} or the notation should be clarified.
  3. [Section 3.2, Eq. (7)] Equation (7) does not state the axis along which the 1D max-pooling is applied; this axis should be specified explicitly so that the reader can verify the correspondence with the recurrence.
  4. [Section 3.3, Eq. (8)] The symbol W is used both for the linear transformation parameter in Section 3.2 and for the logical disjunction approximation in Equation (8); using distinct symbols would avoid confusion.
  5. [Section 4.1, anonymous link] There is a typo in the anonymous repository URL ('Neurlized' should likely be 'Neuralized'), and the link should be checked for correctness.
  6. [Section 4.3, Table 1] The table caption mentions 'redder colors' but the table is printed without a color legend; the text should either describe the color scale or the table should be readable without color.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RNN is assembled from the externally given DTW dynamic program; the unproven length-shortening approximation and the Eq. 5 index error are correctness risks, not circular reductions.

full rationale

The paper's derivation chain starts from the standard DTW dynamic program (Eq. 2), an external mathematical fact, and then proposes a modified recurrence (Eq. 3) in which the horizontal move is dropped. The RNN in Eqs. 4-7 is a faithful implementation of that modified recurrence (modulo an indexing error in Eq. 5, which writes ht-1[i-1,j] where the recurrence requires ht-1[i,j-1]). Nothing in the construction fits the target labels into the recurrence and then calls that fit a prediction; DTW distances and RNN hidden states are computed from inputs and prototypes, not from supervised targets. The length-shortening step is justified only by a toy illustration and the assertion that distances remain nearly unchanged, so the central 'equivalence to DTW' claim is under-supported and the printed Eq. 5 does not implement the stated recurrence; these are validity and implementation defects, not circularity. The only self-citation is Ref. [3] in the related-work paragraph on data scarcity, and it is not load-bearing. Empirical claims are checked against held-out UCR test splits, so they are externally falsifiable rather than true by construction. No prediction or derivation reduces to its own input, and no self-citation chain forces the conclusion.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The derivation rests on two grid-searched hyperparameters (shortening ratio, prototype count) and on an unproven domain assumption that coordinate merging preserves DTW distance. No new physical or mathematical entities are introduced; the prototype tensor is a learnable model parameter, not an invented extra entity.

free parameters (2)
  • sequence shortening ratio = grid-searched over {0.3, 0.5, 0.8, 0.9}
    Controls how aggressively the length-shortening algorithm compresses prototype sequences; central to the claim that horizontal DTW moves can be dropped; chosen by grid search on each dataset without a stated validation split (Section 4.1).
  • number of prototypes per class = grid-searched over {5, 10, 15, 20}
    Determines the size of the prototype tensor P and the aggregation width; selected per dataset by grid search (Section 4.1).
assumptions (4)
  • standard math The DTW dynamic programming recurrence (Eq. 2) computes the minimum cumulative cost over valid warping paths.
    Standard background result for DTW, invoked in Section 2.
  • domain assumption Merging closest successive coordinates in a prototype sequence preserves DTW alignment distance closely enough that one-to-many horizontal alignments can be eliminated.
    Invoked in Section 3.1 to justify dropping the horizontal move from the DTW recurrence (Eq. 2 to Eq. 3); supported only by two toy schematics in Fig. 1, no theorem or worst-case bound.
  • domain assumption The recurrence with only down and diagonal moves (Eq. 3) is an acceptable proxy for the full DTW distance in classification.
    Underlies the 'equivalent RNN' claim; the paper explicitly says 'approximately optimized' but the abstract and conclusions call it equivalent.
  • domain assumption Softmax-normalized DTW scores can be aggregated by an approximate logical disjunction to form class probabilities.
    Equation 8 follows prior work by Li and Srikumar [17]; treated as valid without empirical validation in this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bridging Neural Networks and Dynamic Time Warping for Adaptive Time Series Classification." pith.science (2026). https://pith.science/paper/CE4YYCKN

@misc{pith2026250709826,
  author       = {Pith},
  title        = {Pith review of: Bridging Neural Networks and Dynamic Time Warping for Adaptive Time Series Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CE4YYCKN}},
  note         = {Machine review of arXiv:2507.09826}
}
read the original abstract

Neural networks have achieved remarkable success in time series classification, but their reliance on large amounts of labeled data for training limits their applicability in cold-start scenarios. Moreover, they lack interpretability, reducing transparency in decision-making. In contrast, dynamic time warping (DTW) combined with a nearest neighbor classifier is widely used for its effectiveness in limited-data settings and its inherent interpretability. However, as a non-parametric method, it is not trainable and cannot leverage large amounts of labeled data, making it less effective than neural networks in rich-resource scenarios. In this work, we aim to develop a versatile model that adapts to cold-start conditions and becomes trainable with labeled data, while maintaining interpretability. We propose a dynamic length-shortening algorithm that transforms time series into prototypes while preserving key structural patterns, thereby enabling the reformulation of the DTW recurrence relation into an equivalent recurrent neural network. Based on this, we construct a trainable model that mimics DTW's alignment behavior. As a neural network, it becomes trainable when sufficient labeled data is available, while still retaining DTW's inherent interpretability. We apply the model to several benchmark time series classification tasks and observe that it significantly outperforms previous approaches in low-resource settings and remains competitive in rich-resource settings.

Figures

Figures reproduced from arXiv: 2507.09826 by the authors.

Figure 1
Figure 1. Illustration of the dynamic length-shortening process. (a) The original align￾ment, where the bottom sequence contains redundant consecutive coordinates, (b) the alignment after adaptive length shortening, where redundant coordinates are merged. Furthermore, as illustrated by the green-marked correspondences in [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the computation flow in the proposed neural model. Based on the aforementioned model definition, the overall computational flow is illustrated in Fig.2. Referring to the equivalent transformation in Equation (5), this computation can be decomposed into following two parts: (1) Computation of ∆t(S). At time step t, the model takes the sample’s feature representation xt ∈ R D as input, along with the h… view at source ↗
Figure 3
Figure 3. Accuracy comparison between our model and various baselines. Each point represents a dataset, with the x-axis showing baseline accuracy and the y-axis showing our model’s accuracy. Red points indicate datasets where distance-based methods are preferred. Points above the diagonal represent cases where our model outperforms the baseline. We compare our model with various baseline methods to verify its effective￾ness i… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 28 canonical work pages

  1. [1]

    Data mining and knowledge discovery31, 606–660 (2017)

    Bagnall, A., Lines, J., Bostrom, A., Large, J., Keogh, E.: The great time series classification bake off: a review and experimental evaluation of recent algorithmic advances. Data mining and knowledge discovery31, 606–660 (2017)

    work page 2017

  2. [2]

    In: EPIA Conference on Artificial Intelligence

    Cerqueira, V., Moniz, N., Inácio, R., Soares, C.: Time series data augmentation as an imbalanced learning problem. In: EPIA Conference on Artificial Intelligence. pp. 335–346. Springer (2024)

    work page 2024

  3. [3]

    ACM SIGKDD explorations newsletter26(1), 34–48 (2024) Bridging NNs and DTW for Adaptive Time Series Classification 15

    Chu, Z., Wang, Z., Zhang, W.: Fairness in large language models: A taxonomic survey. ACM SIGKDD explorations newsletter26(1), 34–48 (2024) Bridging NNs and DTW for Adaptive Time Series Classification 15

    work page 2024

  4. [4]

    https://doi.org/10.48550/ARXIV.1810.07758, https://arxiv.org/abs/1810.07758

    Dau, H.A., Bagnall, A., Kamgar, K., Yeh, C.C.M., Zhu, Y., Gharghabi, S., Ratanamahatana, C.A., Keogh, E.: The ucr time series archive (2018). https://doi.org/10.48550/ARXIV.1810.07758, https://arxiv.org/abs/1810.07758

    arXiv doi:10.48550/arxiv.1810.07758 2018

  5. [5]

    Data Mining and Knowledge Discovery 34(5), 1454–1495 (2020)

    Dempster, A., Petitjean, F., Webb, G.I.: Rocket: exceptionally fast and accurate time series classification using random convolutional kernels. Data Mining and Knowledge Discovery 34(5), 1454–1495 (2020)

    work page 2020

  6. [6]

    In: The Twelfth International Conference on Learning Representations (2024)

    Early, J., Cheung, G., Cutajar, K., Xie, H., Kandola, J., Twomey, N.: Inherently in- terpretable time series classification via multiple instance learning. In: The Twelfth International Conference on Learning Representations (2024)

    work page 2024

  7. [7]

    Neural computation 9(8), 1735–1780 (1997)

    Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural computation 9(8), 1735–1780 (1997)

    1997

  8. [9]

    Data Mining and Knowledge Discovery34(6), 1936–1962 (2020)

    Ismail Fawaz, H., Lucas, B., Forestier, G., Pelletier, C., Schmidt, D.F., Weber, J., Webb, G.I., Idoumghar, L., Muller, P.A., Petitjean, F.: Inceptiontime: Finding alexnet for time series classification. Data Mining and Knowledge Discovery34(6), 1936–1962 (2020)

    work page 2020

Show all 36 references
  1. [10]

    Pattern recognition44(9), 2231–2240 (2011)

    Jeong, Y.S., Jeong, M.K., Omitaomu, O.A.: Weighted dynamic time warping for time series classification. Pattern recognition44(9), 2231–2240 (2011)

    2011

  2. [11]

    In: Zhou, Z.H

    Jia, Z., Lin, Y., Wang, J., Wang, X., Xie, P., Zhang, Y.: Salientsleepnet: Mul- timodal salient wave detection network for sleep staging. In: Zhou, Z.H. (ed.) Proceedings of the Thirtieth International Joint Conference on Artificial In- telligence, IJCAI-21. pp. 2614–2620. Int...

    2021 doi

  3. [12]

    In: International Conference on Case-Based Reasoning

    Keane, M.T., Kenny, E.M.: How case-based reasoning explains neural networks: A theoretical analysis of xai using post-hoc explanation-by-example from a survey of ann-cbr twin-systems. In: International Conference on Case-Based Reasoning. pp. 155–171. Springer (2019)

    2019

  4. [13]

    In: Twenty-Eighth International Joint Conferences on Artifical Intel- ligence (IJCAI), Macao, 10-16 August 2019

    Kenny,E.M.,Keane,M.T.:Twin-systemstoexplainartificialneuralnetworksusing case-based reasoning: Comparative tests of feature-weighting methods in ann-cbr twins for xai. In: Twenty-Eighth International Joint Conferences on Artifical Intel- ligence (IJCAI), Macao, 10-16 August 20...

    2019

  5. [14]

    arXiv preprint arXiv:2205.12729 (2022)

    Kook, L., Götschi, A., Baumann, P.F., Hothorn, T., Sick, B.: Deep interpretable ensembles. arXiv preprint arXiv:2205.12729 (2022)

    2022 arXiv

  6. [15]

    In: International Conference on Big Data Analytics and Knowledge Discovery

    Labaien, J., Zugasti, E., Carlos, X.D.: Contrastive explanations for a deep learning model on time-series data. In: International Conference on Big Data Analytics and Knowledge Discovery. pp. 235–244. Springer (2020)

    2020

  7. [16]

    In: The Thirty-Third International Flairs Conference (2020)

    Leonardi, G., Montani, S., Striani, M.: Deep feature extraction for representing and classifying time series cases: towards an interpretable approach in haemodialysis. In: The Thirty-Third International Flairs Conference (2020)

    2020

  8. [17]

    In: Pro- ceedings of the 57th Annual Meeting of the Association for Computational Linguis- tics

    Li, T., Srikumar, V.: Augmenting neural networks with first-order logic. In: Pro- ceedings of the 57th Annual Meeting of the Association for Computational Linguis- tics. pp. 292–302. Association for Computational Linguistics, Florence, Italy (Jul 2019). https://doi.org/10.1865...

    2019 doi

  9. [18]

    Data Mining and Knowledge Discovery29, 565–592 (2015)

    Lines, J., Bagnall, A.: Time series classification with ensembles of elastic distance measures. Data Mining and Knowledge Discovery29, 565–592 (2015)

    2015

  10. [19]

    arXiv preprint arXiv:1801.00631 (2018) 16 J

    Marcus, G.: Deep learning: A critical appraisal. arXiv preprint arXiv:1801.00631 (2018) 16 J. Qu et al

    2018 arXiv

  11. [20]

    IEEE transactions on pattern analysis and machine intelligence31(2), 306–318 (2008)

    Marteau, P.F.: Time warp edit distance with stiffness adjustment for time series matching. IEEE transactions on pattern analysis and machine intelligence31(2), 306–318 (2008)

    2008

  12. [21]

    Scientific Reports15(1), 6932 (2025)

    Rolando,M.,Raggio,V.,Naya,H.,Spangenberg,L.,Cagnina,L.:Alabeledmedical records corpus for the timely detection of rare diseases using machine learning approaches. Scientific Reports15(1), 6932 (2025)

    2025

  13. [22]

    Data Mining and Knowledge Discovery35(2), 401– 449 (2021)

    Ruiz, A.P., Flynn, M., Large, J., Middlehurst, M., Bagnall, A.: The great multi- variate time series classification bake off: a review and experimental evaluation of recent algorithmic advances. Data Mining and Knowledge Discovery35(2), 401– 449 (2021)

    2021

  14. [23]

    CEUR Workshop Proceedings (2017)

    Sani, S., Wiratunga, N., Massie, S.: Learning deep features for knn-based human activity recognition. CEUR Workshop Proceedings (2017)

    2017

  15. [24]

    Data Mining and Knowledge Discovery29, 1505–1530 (2015)

    Schäfer, P.: The boss is concerned with time series classification in the presence of noise. Data Mining and Knowledge Discovery29, 1505–1530 (2015)

    2015

  16. [25]

    In: Proceedings of the 31st International Conference on Neural Information Processing Systems

    Snell, J., Swersky, K., Zemel, R.: Prototypical networks for few-shot learning. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. p. 4080–4090. NIPS’17, Curran Associates Inc., Red Hook, NY, USA (2017)

    2017

  17. [26]

    IEEE transactions on Knowledge and Data Engineering25(6), 1425–1438 (2012)

    Stefan, A., Athitsos, V., Das, G.: The move-split-merge metric for time series. IEEE transactions on Knowledge and Data Engineering25(6), 1425–1438 (2012)

    2012

  18. [27]

    In: 2021 Interna- tional Joint Conference on Neural Networks (IJCNN)

    Sun, J., Takeuchi, S., Yamasaki, I.: Prototypical inception network with cross branch attention for time series classification. In: 2021 Interna- tional Joint Conference on Neural Networks (IJCNN). pp. 1–7 (2021). https://doi.org/10.1109/IJCNN52387.2021.9533440

    2021

  19. [28]

    In: Proceedings of the Thirty- First AAAI Conference on Artificial Intelligence

    Szegedy, C., Ioffe, S., Vanhoucke, V., Alemi, A.A.: Inception-v4, inception-resnet and the impact of residual connections on learning. In: Proceedings of the Thirty- First AAAI Conference on Artificial Intelligence. p. 4278–4284. AAAI’17, AAAI Press (2017)

    2017

  20. [29]

    Tan, Y., Yang, C., Wei, X., Chen, C., Liu, W., Li, L., Zhou, J., Zheng, X.: Metacare++: Meta-learning with hierarchical subtyping for cold- start diagnosis prediction in healthcare data. pp. 449–459 (07 2022). https://doi.org/10.1145/3477495.3532020

    2022

  21. [30]

    The VLDB Journal15, 1–20 (2006)

    Vlachos, M., Hadjieleftheriou, M., Gunopulos, D., Keogh, E.: Indexing multidi- mensional time-series. The VLDB Journal15, 1–20 (2006)

    2006

  22. [31]

    Knowledge-Based Systems 243, 108459 (2022)

    Wang, T., Liu, Z., Zhang, T., Hussain, S.F., Waqas, M., Li, Y.: Adaptive feature fusion for time series classification. Knowledge-Based Systems 243, 108459 (2022). https://doi.org/https://doi.org/10.1016/j.knosys.2022.108459, https://www.sciencedirect.com/science/article/pii/S...

    2022

  23. [32]

    arXiv preprint arXiv:1906.00917 (2019)

    Wang, Y., Emonet, R., Fromont, E., Malinowski, S., Menager, E., Mosser, L., Tavenard, R.: Learning interpretable shapelets for time series classification through adversarial regularization. arXiv preprint arXiv:1906.00917 (2019)

    2019 arXiv

  24. [33]

    Wang, Y.: Designing Deep Methods to Improve Machine Learning Interpretability. Ph.D. thesis, Washington State University (2022)

    2022

  25. [34]

    In: The Thirteenth International Conference on Learning Representations

    Wen, Y., Ma, T., Luss, R., Bhattacharjya, D., Fokoue, A., Julius, A.A.: Shedding light on time series classification using interpretability gated networks. In: The Thirteenth International Conference on Learning Representations

  26. [35]

    IEEE Sensors Journal (2024) Bridging NNs and DTW for Adaptive Time Series Classification 17

    Zhang, B., Li, L., Liang, G., Tan, C., Dong, F.: Elastic slow feature prototypical network for few-shot fault diagnosis of industrial processes. IEEE Sensors Journal (2024) Bridging NNs and DTW for Adaptive Time Series Classification 17

    2024

  27. [36]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Zhang, X., Gao, Y., Lin, J., Lu, C.T.: Tapnet: Multivariate time series classification with attentional prototypical network. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 34, pp. 6845–6852 (2020)

    2020

  28. [37]

    In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S

    Zhou, X., Liu, X., Zhai, D., Jiang, J., Gao, X., Ji, X.: Prototype-anchored learning for learning with imperfect annotations. In: Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., Sabato, S. (eds.) Proceedings of the 39th International Conference on Machine Learni...

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.