REVIEW 3 major objections 7 minor 19 references
Learning Soft Sparse Shapes for Efficient Time-Series Classification
T0 review · 3 major / 7 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read Merging weak shapes, not deleting them, tops 128 UCR benchmarks
desk verdict Soft sparsification of shapelets is a real idea, but the SOTA claim rests on a comparison protocol that flatters it. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the soft shape sparsification of Eqs. (3)-(5). A parameter-shared gated attention head, $\alpha(\hat{S}^m_{n,p}) = \sigma(W_2 \tanh(W_1 \hat{S}^m_{n,p} + b_1) + b_2)$, scores every subsequence embedding by its estimated classification contribution; the top-$\eta$ proportion become scaled soft shapes $\alpha \hat{S}$, and the rest are merged into a single fused embedding $e^S_{n,\mathrm{fused}} = \sum_{p \in E} \alpha(\hat{S}^m_{n,p}) \hat{S}^m_{n,p}$. This merge-not-delete operation is what lets the model keep all subsequence information while shrinking its input from $J$ shapes to $\eta J + 1$. The second mechanism is the soft shape learning block: a mixture-of-experts router with softmax gating and TOP-$k$ selection (here $k = 1$, Eqs. (6)-(9)) routes each soft shape to a class-specific expert for intra-shape pattern learning, while a shared Inception-style 1D convolutional expert on the reshaped sequence of shapes (Eq. (12)) captures inter-shape temporal dependencies. Importance and load-balancing losses (Eqs. (10)-(11)) keep the experts from collapsing onto a few favored routes.
What would settle it
Two checks would settle the claim. First, replace the fused low-scoring shape embedding of Eq. (5) with a constant vector on the 18-dataset benchmark: if accuracy does not drop, the merge-retains-information mechanism is not doing the work claimed. Second, rerun the re-implemented baselines (TSLANet, ModernTCN, ShapeConv) on the original UCR train/test split with their published hyperparameters: if they recover their published accuracies, SoftShape's reported margin of 0.0129 in average accuracy has not survived contact with the standard protocol.
Extended reading notes
Core claim
SoftShape's central claim is that sparsification of shapelet candidates should be soft, not hard. Instead of selecting the top shapes and discarding the rest — the move shared by classical shapelet transforms and recent deep shapelet layers — the model scales the top-$\eta$ proportion of subsequence embeddings by their attention scores and fuses the remaining embeddings into a single weighted-sum shape appended to the sequence. Every subsequence therefore survives sparsification in some form, while the sequence shrinks from $J$ shapes to $\eta J + 1$. On this sparsified input, a router activates exactly one class-specific expert per shape to capture intra-shape patterns, and a shared convolutional expert treats the shapes as a sequence to capture inter-shape temporal dependencies. The paper reports this design achieves average test accuracy 0.9334 and average rank 2.72 across 128 UCR datasets, ahead of the best compared deep baseline TSLANet (0.9205, rank 3.68) with a Wilcoxon p-value of $1.06 \times 10^{-3}$, and that the learned attention scores highlight the same discriminative regions a human would pick out.
Load-bearing premise
The evaluation rests on 19 baselines, several re-implemented or borrowed from another study, all trained under the paper's own 60/20/20 re-split of the UCR data; if those baselines were tuned below their published best, the reported accuracy edge would shrink or vanish.
Editorial extensions
If this is right
- A shapelet-based deep model can sparsify without losing information: merging low-scoring subsequences into a fused shape keeps their contribution while cutting the number of shapes the learner processes from $J$ to $\eta J + 1$.
- Interpretability comes from the same scores that drive sparsification: the multiple-instance-learning visualizations on Trace and Lightning2 show that high-scoring shapes are exactly the regions where the classes differ sharply.
- Efficiency tracks the sparsity ratio: at $\eta = 0.5$ the shared expert reads roughly half the shapes, and on long sequences SoftShape trains faster than the attention-based and transformer baselines reported.
- Fusing up to 50% of shapes causes no significant accuracy drop while 90% does (Table 3), which directly supports the paper's claim that hard shapelet methods lose critical patterns by discarding too much.
- The same model transfers to forecasting without hyperparameter adjustment, beating TS2Vec and TimesNet on four ETT datasets (Appendix B.5), so the soft-shape machinery is not specific to classification.
Reading between the lines
- The fused shape is effectively a learned, class-contribution-weighted summary of a series' non-discriminative background; a testable extension would be learning several fused shapes per series instead of one, which might preserve more information at sparsity ratios above 50%.
- The comparison set excludes the strongest published UCR classifiers such as HIVE-COTE 2.0 and baselines tuned to their published accuracies, so how large the real-world margin is under the standard UCR protocol remains open.
- Read against Table 2, removing the inter-shape module costs more (0.9334 to 0.9022) than removing soft sparsification (to 0.9123), suggesting a substantial share of the reported gain may come from the shape-sequence CNN and MoE blocks rather than from the merge-not-delete mechanism alone; testing a plain patch-based model with the same learning blocks on the same protocol would size that share.
- The forecasting results hint that soft sparsification could serve as a general tokenizer for time series; applying the same pipeline to anomaly detection or long-horizon forecasting, where rare local patterns carry the signal, would be a direct test.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes SoftShape, a deep-learning architecture for univariate time series classification. The model converts an input series into overlapping subsequence ('shape') embeddings via a 1D CNN, scores each embedding with a gated attention head, rescales the top-η fraction of shapes by their scores, and fuses all remaining low-scoring shapes into a single weighted-sum embedding ('soft shape sparsification'). The sparsified shapes then pass through a soft shape learning block that combines a mixture-of-experts router for intra-shape pattern learning with a shared Inception-style expert for inter-shape temporal modeling. Training is end-to-end with cross-entropy plus two MoE load-balancing losses. The authors evaluate on 128 UCR datasets under a merged 60/20/20 five-fold cross-validation protocol, reporting an average accuracy of 0.9334 and an average rank of 2.72 against 17 baselines, with Wilcoxon signed-rank p-values; they also provide ablations, hyperparameter studies, runtime measurements, and attention visualizations. The central claim is that SoftShape outperforms state-of-the-art TSC methods while remaining interpretable.
Significance. If the empirical claim is substantiated, the result is significant: it suggests that a shapelet-inspired architecture with soft sparsification and MoE-based intra- and inter-shape learning can outperform the strongest compared convolutional and Transformer baselines on a large benchmark while offering attention-based explanations. The model is clearly specified and the forward-pass mathematics are coherent. Strengths of the manuscript include the breadth of the main evaluation (128 UCR datasets), the ablation battery, the hyperparameter sensitivity analysis, the release of code, and the explicit reporting of Wilcoxon p-values. The main unresolved question is whether the comparison protocol is fair to the baselines and whether the reported margins would persist under better-tuned baselines and with variance information; the evidence currently presented is not sufficient to fully support the headline state-of-the-art claim.
major comments (3)
- [Appendix A.2-A.3, Table 9] The fairness of the baseline comparison under the merged 60/20/20 protocol is not established. The same optimizer, learning rate, batch size, and early-stopping rule are applied to all deep baselines without per-method hyperparameter tuning, and several re-implemented baselines score far below their published accuracy on the same UCR datasets: e.g., AllGestureWiimoteX ShapeConv 0.317 and ElectricDevices ModernTCN 0.350, with average accuracies of 0.7688 and 0.7938, respectively. The absence of HIVE-COTE 2.0, the strongest prior UCR classifier in recent bake-offs, further weakens the state-of-the-art claim. Because SoftShape's margin over the strongest compared baseline TSLANet is only 0.0129 in average accuracy, the conclusion that SoftShape 'outperforms state-of-the-art methods' is load-bearing on the quality of these baseline runs; if the baselines were undertuned, the margin could shrink or disappear. Please provide per-method hyperparameter tuning (or justification that the common setting is appropriate for each baseline), report published-level reference numbers on the official UCR splits, and add or justify the omission of HIVE-COTE 2.0.
- [Section 5.3 and Tables 3, 5, 6, 13, 15-21] Several key hyperparameters of SoftShape (η=0.5, q=4, L=2, k=1, λ=0.001) are selected on an 18-dataset subset of the same 128-dataset benchmark, and the main results in Table 1 include those same 18 datasets. This makes the comparison partially in-sample for those datasets, since the model configuration was chosen after observing their test accuracy, and the reported Wilcoxon p-values do not account for this selection. To support a state-of-the-art claim, the authors should either fix all hyperparameters before evaluating on UCR, exclude the 18 development datasets from the main comparison, or use a separate validation benchmark. At minimum, the paper should explicitly state that the main results are not fully held-out with respect to hyperparameter selection.
- [Appendix A.3, Implementation Details] The implementation description is internally contradictory about the number of seeds. The text states both 'we calculate the average test accuracy using five-fold test sets from a single seed' and 'Each experiment is conducted five times with five different random seeds'; no standard deviations or error bars are reported for any of the headline numbers in Table 1. Because the primary evidence is a small accuracy margin (0.9334 vs. 0.9205) and the claimed statistical significance is based on paired tests, the paper must clarify the seed protocol and report variance measures (e.g., per-dataset standard deviation across seeds or folds) for SoftShape and at least the strongest baselines. Without this, the reader cannot assess whether the reported difference is robust to seed variation.
minor comments (7)
- [Section 2.2] The phrase 'improving model interop-erability' appears to be a typo for 'interpretability'.
- [Algorithm 1] In Step 3, 'residul way' should be 'residual way'.
- [Section 4.4, Eq. (13)] The summation index in the conjunctive pooling equation runs from i=0 to Num, which would produce Num+1 terms while the normalization factor is 1/Num; please check whether the intended range is i=1,...,Num or i=0,...,Num-1.
- [Table 1] The 'Win' column sums to more than the number of datasets because ties are apparently counted for multiple methods; the caption should state how ties are handled.
- [Section 5.1] The Wilcoxon p-values are reported for each baseline without correction for multiple comparisons across 17 baselines; please state whether any correction was applied or note that the p-values are uncorrected.
- [Appendix B.5] The time series forecasting experiments (Tables 22 and 23) are not part of the paper's stated contributions and show SoftShape underperforming iTransformer; if kept, they should be better motivated and their limitations acknowledged, or they should be removed.
- [Figure 2 caption] The tensor dimension annotations such as [B,D,5] and [5B,1,D] are not explained in the caption; adding a short explanation of the reshape operations would improve readability.
Circularity Check
No significant circularity: SoftShape's test accuracy is an out-of-sample evaluation; only minor self-citation of baseline numbers appears, and no derivation reduces to its own inputs.
full rationale
The paper's central claim is an empirical accuracy comparison, not a derived quantity. SoftShape is trained with cross-entropy on 60% training folds and evaluated on 20% held-out test folds per the 3:1:1 split described in Appendix A.1 and A.3, so the reported test accuracy is an out-of-sample evaluation rather than a fitted parameter renamed as a prediction. The attention scores in the soft shape sparsification (Eqs. 3-5) are learned end-to-end and are not defined in terms of test accuracy; the final classifier uses conjunctive pooling (Eq. 13) but this is a model architecture, not a circular reduction. The hyperparameters eta, q, L, k, and lambda are selected on 18 UCR datasets and then fixed for the main 128-dataset experiments; this is a tuning and potential selection-on-test concern, but the main table results are still computed on held-out folds under fixed settings, so it does not make the accuracy claim equivalent to the tuning objective. The only notable self-citation is the reuse of nine baseline columns from Ma et al. (2024), a prior survey with overlapping authors, and the adoption of the same split protocol from that source. Those are empirical inputs to the comparison rather than derivations of SoftShape's result, and the principal margin against TSLANet, InceptionTime, MR-H, and RDST is not imported from that source. Therefore, no load-bearing step reduces to its own input, and the appropriate finding is no significant circularity, with a score of 1 reflecting the minor self-citation of baseline numbers rather than any circular derivation.
Assumptions & free parameters
free parameters (7)
- eta (sparsify ratio) =
0.5
- k (MoE activated experts) =
1
- q (sliding step size) =
4
- L (model depth) =
2
- lambda (MoE loss weight) =
0.001
- d_attn (attention hidden dim) =
8
- warmup epochs =
150
assumptions (4)
- domain assumption Gated attention with independent shape embeddings (Ilse et al., 2018) captures the class-discriminative power of each subsequence.
- domain assumption The 18 selected UCR datasets (Appendix B.3) are representative for choosing eta, k, q, L for the full 128-dataset benchmark.
- ad hoc to paper Fusing low-scoring shape embeddings into a single weighted sum (Eq. 5) preserves non-trivial information rather than destroying it.
- ad hoc to paper Exponential of gating probabilities in Eq. 9 is a valid scaling for the MoE expert output.
Cite this review
Pith. "Pith review of Learning Soft Sparse Shapes for Efficient Time-Series Classification." pith.science (2026). https://pith.science/paper/Y3WM4PBW
@misc{pith2026250506892,
author = {Pith},
title = {Pith review of: Learning Soft Sparse Shapes for Efficient Time-Series Classification},
year = {2026},
howpublished = {\url{https://pith.science/paper/Y3WM4PBW}},
note = {Machine review of arXiv:2505.06892}
}
read the original abstract
Shapelets are discriminative subsequences (or shapes) with high interpretability in time series classification. Due to the time-intensive nature of shapelet discovery, existing shapelet-based methods mainly focus on selecting discriminative shapes while discarding others to achieve candidate subsequence sparsification. However, this approach may exclude beneficial shapes and overlook the varying contributions of shapelets to classification performance. To this end, we propose a Soft sparse Shapes (SoftShape) model for efficient time series classification. Our approach mainly introduces soft shape sparsification and soft shape learning blocks. The former transforms shapes into soft representations based on classification contribution scores, merging lower-scored ones into a single shape to retain and differentiate all subsequence information. The latter facilitates intra- and inter-shape temporal pattern learning, improving model efficiency by using sparsified soft shapes as inputs. Specifically, we employ a learnable router to activate a subset of class-specific expert networks for intra-shape pattern learning. Meanwhile, a shared expert network learns inter-shape patterns by converting sparsified shapes into sequences. Extensive experiments show that SoftShape outperforms state-of-the-art methods and produces interpretable results.
Figures
Figures from the paper (7 more)
Reference graph
Works this paper leans on
-
[1]
To address this, Grabocka et al
Directly using subsequences of varying lengths as model inputs significantly increases computational complexity and requires additional preprocessing to standardize input dimensions. To address this, Grabocka et al. (2014); Qu et al. (2024) employ a validation set to select a fixed-length m of subsequence for shapelet learning, where m is defined as γT , ...
work page 2014
-
[3]
Crafting papers on machine learning
Langley, P. Crafting papers on machine learning. In Langley, P. (ed.),Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp. 1207–1216, Stan- ford, CA,
2000
-
[4]
The rateη in Equation (5) is set to 50%
Thedattn in Equation (3) is set to 8 (Early et al., 2024). The rateη in Equation (5) is set to 50%. To mitigate the issue of inaccurate attention score evaluations for most shapes 14https://openreview.net/attachment?id=O8ouVV8PjF&name=supplementary material 15https://github.com/xuanmay2701/shapeformer 16https://github.com/luodhhh/ModernTCN 17https://githu...
work page 2024
-
[5]
Rakthanmanon, T. and Keogh, E. Fast shapelets: A scalable algorithm for discovering time series shapelets. In pro- ceedings of the 2013 SIAM International Conference on Data Mining, pp. 668–676. SIAM,
work page 2013
-
[6]
Time series classification from scratch with deep neural networks: A strong base- line
Wang, Z., Yan, W., and Oates, T. Time series classification from scratch with deep neural networks: A strong base- line. In 2017 International joint conference on neural networks (IJCNN), pp. 1578–1585,
work page 2017
-
[8]
12 Learning Soft Sparse Shapes for Efficient Time-Series Classification A. Experimental Setup A.1. Datasets In this study, we employ the UCR time series archive (Dau et al., 2019), a widely recognized benchmark for time series classification tasks (Ismail Fawaz et al., 2019; Middlehurst et al., 2024), to systematically evaluate the effectiveness of the pr...
work page 2019
-
[9]
and the recommendations of (Dau et al., 2019), we integrate the raw training and test sets and employ a five-fold cross-validation strategy to partition the dataset into training, validation, and test sets in a 3:1:1 ratio. Consistent with the approach in (Ma et al., 2024), each fold is sequentially designated as the test set, while the remaining four fol...
work page 2024
-
[10]
Train” represents the count of samples within the raw training set. “Test
is a general framework for time series representation learning across various semantic levels, aimed at tasks such as time series classification, forecasting, and anomaly detection. In the context of the time series classification task, the authors utilized temporal convolutional networks (TCN) as the foundational framework, leveraging the low-dimensional...
work page 1965
Show all 19 references
-
[12]
demonstrate that HIVE-COTE version 2 (HC2) (Middlehurst et al., 2021), which is a hybrid-based method, MultiROCKET-Hydra (MR-H) (Dempster et al., 2023), a convolution-based method, and RDST, a shapelet-based approach, exhibit strong classification performance. Given that HC2 i...
2021
-
[16]
Figure 6 illustrates the critical difference diagram and significance analysis results for SoftShape and the 17 baseline methods on the UCR 128 time series dataset
Specifically, the test classification accuracy results of nine methods, including FCN, T-Loss, SelfTime, TS-TCC, TST, TS2Vec, TimesNet, PatchTST, and GPT4TS, are obtained from (Ma et al., 2024). Figure 6 illustrates the critical difference diagram and significance analysis res...
2024
-
[21]
B.5. Time Series Forecasting Results For the forecasting task, we evaluate SoftShape on the ETTh1, ETTh2, ETTm1, and ETTm2 datasets, using TS2Vec (Yue et al., 2022), TimesNet (Wu et al., 2023a), PatchTST (Nie et al., 2023), GPT4TS (Zhou et al., 2023), and iTransformer (Liu et ...
2022
-
[23]
The results demonstrate that SoftShape outperforms TS2Vec and TimesNet, highlighting its potential for time series forecasting tasks
Notably, we do not adjust hyperparameters for SoftShape across the four datasets, nor do we modify its shape embedding layer as done in (Nie et al., 2023). The results demonstrate that SoftShape outperforms TS2Vec and TimesNet, highlighting its potential for time series foreca...
2023
-
[25]
Among these, w/o II refers to the w/o Intra & Inter method
Besides, the selected datasets cover diverse application areas, including handwritten font sequence recognition (e.g., the WordSynonymsdataset), human activity recognition (e.g., the CricketX dataset), and medical diagnosis 21 Learning Soft Sparse Shapes for Efficient Time-Ser...
-
[500]
For all UCR datasets, we apply a uniform normalization strategy to standardize each time series within the dataset (Ismail Fawaz et al., 2019)
Also, for all baseline method training, we implement a consistent early stopping strategy based on validation set loss values (Ma et al., 2024). For all UCR datasets, we apply a uniform normalization strategy to standardize each time series within the dataset (Ismail Fawaz et ...
2024
-
[2019]
F., Weber, J., Webb, G
Ismail Fawaz, H., Lucas, B., Forestier, G., Pelletier, C., Schmidt, D. F., Weber, J., Webb, G. I., Idoumghar, L., Muller, P.-A., and Petitjean, F. Inceptiontime: Finding alexnet for time series classification. Data Mining and Knowledge Discovery, 34(6):1936–1962,
1936
-
[2021]
Moirai-moe: Empowering time series foundation models with sparse mixture of experts
Liu, X., Liu, J., Woo, G., Aksu, T., Liang, Y ., Zimmer- mann, R., Liu, C., Savarese, S., Xiong, C., and Sahoo, D. Moirai-moe: Empowering time series foundation models with sparse mixture of experts. arXiv preprint arXiv:2410.10469, 2024a. Liu, Y ., Hu, T., Zhang, H., Wu, H., ...
-
[2022]
St-moe: Designing stable and transferable sparse expert models
Zoph, B., Bello, I., Kumar, S., Du, N., Huang, Y ., Dean, J., Shazeer, N., and Fedus, W. St-moe: Designing stable and transferable sparse expert models. arXiv preprint arXiv:2202.08906,
-
[2023]
Furthermore, Middlehurst et al
for time series classification tasks. Furthermore, Middlehurst et al. (2024) conduct a comprehensive analysis of the advantages and limitations of eight categories of time series classification methods: distance-based, feature-based, interval-based, shapelet-based, dictionary-...
2024
-
[2024]
Self-supervised time series representation learning by inter-intra relational reasoning
Fan, H., Zhang, F., and Gao, Y . Self-supervised time series representation learning by inter-intra relational reasoning. arXiv preprint arXiv:2011.13548,
2011 arXiv
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.