Pith. sign in

REVIEW 3 major objections 5 minor 17 references

Making Sense of Touch: Unsupervised Shapelet Learning in Bag-of-words Sense

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read NN-STNE learns unsupervised shapelet features from time series by mapping each series to Student-t kernel distances, and clustering those features beats raw KMeans and UDFS on the reported benchmarks.

desk verdict A plausible but under-supported unsupervised shapelet recipe whose min-pooling step probably loses the multi-event information the paper cares about; worth a peer-review look if the authors add baselines and ablations. read the letter →

arxiv 2502.04167 v1 pith:IV5VUPJ3 submitted 2025-02-06 cs.LG cs.RO

Zhicong Xian , Tabish Chaudhary , Jürgen Bock This is my paper · ORCID
classification cs.LGcs.RO
keywords unsupervisedshapeletlearningt-SNEtimeseriesclusteringtouchsensingforcesignalsroboticsUCRdatasetnormalizedcross-correlation
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 proposes NN-STNE, an unsupervised neural network that learns a small set of time-series sub-sequences, called shapelets, from unlabeled data. Each input time series is reduced to a vector of distances to these shapelets, and a Student-t distribution converts those distances into membership probabilities. Feeding these probability features into KMeans improves clustering accuracy over raw KMeans and over UDFS features on all five tested data sets, including a robot switch-pushing task with force signals. The paper argues that the t-distributed embedding avoids the crowding problem in low-dimensional shapelet space, and that L1 regularization on shapelet weights automatically selects useful shapelet lengths. If the result holds, unsupervised shapelet features provide an interpretable, label-free preprocessing step for time-series clustering, especially for touch and force sensing in robotics.

What carries the argument

The central object is the shapelet membership vector $\mathbf{q}_{i,:} \in \mathbb{R}^K$, computed for each time series by (i) sliding a shapelet of length $M$ across the series, (ii) scoring each window by normalized cross-correlation, (iii) keeping only the window with the smallest distance $F_{i,k} = \min_j D_{i,j,k}$ per shapelet, and (iv) applying a Student-t kernel to these minimal distances to get probabilities. Those probabilities are trained by minimizing $\mathrm{tr}(q^T L_G q) + \lambda \|H\|_2^2 + \beta \sum_{k,l} |s_{k,l}|$, where $L_G$ is the Gaussian-kernel graph Laplacian that preserves local structure, $H$ penalizes pairwise-similar shapelets, and the L1 term prunes shapelet weights to zero so the effective shapelet length is learned. This turns shapelet discovery from a brute-force search into a differentiable feature-extraction layer.

What would settle it

Cluster a synthetic UCR-style set in which each time series contains the same discriminative shapelet twice, with the first occurrence slightly more correlated with a distractor; if NN-STNE's min-pooling keeps only that first occurrence, clustering accuracy should fall toward chance, while a variant that pools over the top few windows stays accurate.

Watch

Extended reading notes

Core claim

NN-STNE learns a dictionary of $K$ shapelets from unlabeled time series by minimizing a Gaussian-Laplacian objective over t-SNE-style membership probabilities, plus a diversity penalty and an L1 length-selection term. For each shapelet, it finds the single most similar sliding window in a time series via normalized cross-correlation and min-pooling, then treats the resulting distance as a coordinate in a lower-dimensional shapelet space. The paper reports that these coordinates, used as KMeans input features, raise the average clustering Rand index from 0.72 (raw KMeans) and 0.70 (UDFS+KMeans) to 0.84 across ECG200, CBF, FACE FOUR, OSU LEAF, and the robot switch-pushing task, with the largest gain on the robot data (0.74 to 1.0).

Load-bearing premise

The method assumes that for each learned pattern, the single most similar segment in a time series contains all of the pattern's useful information, so if the informative event is not the closest match, the representation throws it away.

Editorial extensions

If this is right

  • On the five benchmark sets in Table II, NN-STNE features followed by KMeans raise average clustering accuracy from 0.72 (raw KMeans) and 0.70 (UDFS+KMeans) to 0.84, a 16.7% relative improvement.
  • The learned features are interpretable: each dimension is the similarity of a time series to one learned shapelet, so the shapelets can be inspected to see which force patterns distinguish robot manipulation outcomes.
  • The L1-norm regularization on shapelet weights lets the method shrink unneeded parts of a shapelet to zero, effectively selecting shapelet length automatically.
  • Using a Student-t kernel instead of a Gaussian kernel addresses the crowding problem in the low-dimensional shapelet space, which the paper argues is why the distance-to-shapelet probabilities are more informative.

Reading between the lines

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

  • The min-pooling step is the likeliest bottleneck: pooling only the closest window per shapelet discards information from secondary events, so an attention-weighted pooling over several windows could extend the method to multi-event recordings.
  • The same shapelet-distance representation could be evaluated on classification as well as clustering, since shapelets are classically discriminative patterns; a k-nearest-neighbor or linear classifier on the features would test whether the learned representation preserves label information.
  • Because the method is unsupervised and outputs distances rather than pseudo-labels, it could serve as a general preprocessing layer for any downstream time-series task, including anomaly detection on force/torque streams in robotics.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes NN-STNE, an unsupervised shapelet feature-learning method for time series. The input time series are slid into windows, compared with learnable shapelets via normalized cross-correlation, min-pooled to one scalar per shapelet, and mapped through a Student-t kernel to a probability-like representation. A graph Laplacian term preserves local structure, a diversity term penalizes similar shapelets, and L1 regularization prunes shapelet entries to select their length. The authors evaluate the learned features by feeding them to KMeans and report clustering accuracy on four UCR data sets and one robot switch-pushing data set, claiming an average improvement of 16.7% over raw KMeans and UDFS+KMeans.

Significance. If the empirical claim held, NN-STNE would be a useful unsupervised preprocessing step for tactile and force time-series data, with the attractive property that the learned shapelets are interpretable. The paper also contains a concrete, implementable architecture and an interesting idea of using L1 regularization to shrink shapelet length. However, the current evidence is narrow: only five data sets, no error bars, no comparison with the shapelet baselines named in Section II, and no sensitivity analysis for the many free parameters. In addition, the min-pooling operation in Section IV.A discards all but the best-matching window per shapelet, which is a strong assumption that is neither argued for nor tested. The unsupervised claim is also weakened by Eq. (1), which sets the number of shapelets from the number of classes C. The paper's contribution is therefore promising but not yet established to the standard of its stated claims.

major comments (3)
  1. [Section IV.A, Time Series Similarity Layer and t-SNE Layer] The min-pooling operation F_{i,k} = min_j D_{i,j,k} reduces each time series to a single scalar per shapelet, so all non-closest sliding windows are discarded. If a recording contains multiple relevant events, or if the closest window is not the informative one, the learned representation deletes precisely the information that shapelets are meant to capture. This directly conflicts with the paper's bag-of-words framing and the switch-pushing example in Fig. 1, where several force events may occur in one trajectory. No ablation, synthetic experiment, or theoretical argument is provided to show that the single best window preserves discriminative information for the five data sets or for robot touch data. This is a load-bearing assumption and needs to be either justified experimentally or removed by a pooled / multi-window representation.
  2. [Section IV.A, Eq. (1)] The number of shapelets is set as K = log2[N x (Q-M) x C], where C is explicitly the total number of class labels. Using class-count information in an unsupervised pipeline is a form of label leakage: even if individual labels are not used, the model capacity is chosen from the number of classes, which requires knowing the ground-truth clustering structure. The unsupervised claim of the paper therefore needs either a label-free formula for K or an explicit discussion of why using C does not compromise the unsupervised setting. As written, Eq. (1) is not consistent with the paper's claim of unsupervised shapelet learning.
  3. [Section V.C, Table II] The empirical claim that NN-STNE features 'achieve competitive results compared to other state-of-art feature selection algorithms' is not supported by the experiments as reported. Table II compares only raw KMeans, UDFS+KMeans, and NN-STNE+KMeans; the shapelet baselines k-Shape [7] and unsupervised shapelet learning [13], both named in Section II and Section V, are not evaluated. No error bars, no repeated-run statistics, no hyperparameter values (M, K, sigma, lambda, beta, alpha), and no clustering metric definition are given, although the text refers to Rand Index in [13] while Table II appears to report accuracy. Without these details, the 16.7% average improvement cannot be assessed for statistical or practical significance, and the comparison is incomplete relative to the stated scope.
minor comments (5)
  1. [Abstract and Section IV.A] The abstract says t-SNE is used 'as a hidden layer', but the method only applies a Student-t kernel to shapelet distances (Eq. (4)); this is not the t-SNE embedding algorithm. The wording should be adjusted to avoid overstating the connection.
  2. [Section IV.A, Eq. (4)] Equation (4) has formatting errors and unclear notation: the exponent is not properly typeset, and the text says Fi,k is a distance 'always larger than 0' without explaining the role of alpha. Please re-typeset the equation and define all symbols explicitly.
  3. [Section V.B, Table II] The evaluation metric is inconsistently described: Section V.B lists ACC, NMI, and Rand Index, but the text says 'Rand Index defined in [13]' while Table II is not labeled with the metric. Please state for each reported number which metric is used and cite the source.
  4. [Section IV.A, Eq. (1)] The derivation of K as 'each shapelet as one bit' is not self-contained; it is unclear why the total number of possible representations equals N x (Q-M) x C or why this product should be represented by 2^K shapelets. A more careful explanation is needed.
  5. [General] There are several typographical issues, including 'An use case application', 'sub-subsequence', 'state-of-art', and the notation 'NN-TSNE' in Section V.C vs. 'NN-STNE' elsewhere; also reference [5] is not cited in the main text and Fig. 1 is not referenced explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: NN-STNE features are optimized by an unsupervised objective against raw input and evaluated on external UCR labels; the only label-informed quantity is the hyperparameter K in Eq. (1).

full rationale

The derivation chain starts from raw time series T, forms windowed normalized-cross-correlation distances D (Eqs. 2-3), min-pools to F, converts to t-distributed probabilities q (Eq. 4), and minimizes a Gaussian-kernel Laplacian objective (Eq. 5) plus diversity and L1 terms (Eq. 6). None of these steps uses cluster labels or the evaluation metric; the learned feature matrix F depends only on T, shapelet initializations from KMeans on unlabeled windows, and hyperparameters. The Table II evaluation feeds F to KMeans and compares against known classes, which is an external benchmark rather than a fit to those classes. The only label-dependent quantity is the number of shapelets K chosen in Eq. (1) using the number of classes C; this is a hyperparameter-setting weakness and a mild information leak, but K is not a prediction of cluster assignments, and no equation identifies the final clustering with Eq. (1). The min-pooling choice F_{i,k} = min_j D_{i,j,k} is a lossy and potentially harmful design decision, but it is not a reduction of the output to the input by construction. All cited methods (Grabocka et al., Zhang et al., van der Maaten, Paparrizos) are external; there are no load-bearing self-citations. Therefore no claimed result is equivalent by construction to its inputs, and the paper is self-contained against external benchmarks.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method is unsupervised and does not hide a fitted label inside its derivation. The main extra degrees of freedom are hyperparameters (M, K, sigma, lambda, beta, alpha) and four modeling assumptions, most notably that min-pooling preserves all relevant information and that L1-zeroed weights define a true shapelet length. No new physical or theoretical entities are introduced.

free parameters (6)
  • M, initial shapelet length per data set = unspecified; chosen by human inspection
    Section IV.A: 'By human inspection, we estimate a possible length of shapelets'; all downstream sliding windows depend on M.
  • K, number of shapelets = from Eq (1), K = log2[N*(Q-M)*C]
    Eq (1) sets K using N, Q, M and the number of classes C; the one-shapelet-equals-one-bit model is an arbitrary heuristic.
  • sigma, Gaussian kernel width = not reported
    G_ij = exp(-||t_i - t_j||^2 / sigma^2) in Eq (5) and H_ij in the diversity term; no value or tuning procedure is given.
  • lambda, diversity weight = not reported
    Weight of ||H||_F^2 in Eq (6), chosen by user; no value or search method reported.
  • beta, L1 regularization weight = not reported
    Weight of sum |s_{k,l}| in Eq (6); user-given but unspecified, and directly controls the claimed automatic length selection.
  • alpha, Student-t degrees of freedom = 1
    Set to 1 in Eq (4) following [12]; fixed but influences the probability mapping.
assumptions (4)
  • domain assumption Euclidean distance between raw time series defines the local structure that should be preserved in shapelet space
    Eq (5) uses Gaussian affinity G_ij = exp(-||t_i - t_j||^2 / sigma^2); no normalization or embedding is described before this distance is computed.
  • ad hoc to paper The single best-matching window per shapelet and time series contains all discriminative information
    Min-pooling F_{i,k} = min_j D_{i,j,k} in Section IV.A discards all but one window per shapelet without justification.
  • ad hoc to paper Each shapelet acts as one bit of a code, so K = log2[N*(Q-M)*C] is the right number of shapelets
    Eq (1) in Section IV.A assumes shapelets are independent bits and that class count C is available, both unstated.
  • ad hoc to paper L1-sparsified shapelet weights can be trimmed to reveal an optimal shapelet length
    Section IV.B asserts zeros will not contribute to normalized cross-correlation and that removing zeros yields true length; this conflates weight sparsity with subsequence length and is not demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Making Sense of Touch: Unsupervised Shapelet Learning in Bag-of-words Sense." pith.science (2026). https://pith.science/paper/IV5VUPJ3

@misc{pith2026250204167,
  author       = {Pith},
  title        = {Pith review of: Making Sense of Touch: Unsupervised Shapelet Learning in Bag-of-words Sense},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IV5VUPJ3}},
  note         = {Machine review of arXiv:2502.04167}
}
read the original abstract

This paper introduces NN-STNE, a neural network using t-distributed stochastic neighbor embedding (t-SNE) as a hidden layer to reduce input dimensions by mapping long time-series data into shapelet membership probabilities. A Gaussian kernel-based mean square error preserves local data structure, while K-means initializes shapelet candidates due to the non-convex optimization challenge. Unlike existing methods, our approach uses t-SNE to address crowding in low-dimensional space and applies L1-norm regularization to optimize shapelet length. Evaluations on the UCR dataset and an electrical component manipulation task, like switching on, demonstrate improved clustering accuracy over state-of-the-art feature-learning methods in robotics.

Figures

Figures reproduced from arXiv: 2502.04167 by the authors.

Figure 1
Figure 1. A robot application: switch pushing task (left) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the network architecture A. Layers in NN-STNE Network Initial Estimation of Number and Length of Shapelets For the length of shapelets we adopt the same strategy as in [6]. By human inspection, we estimate a possible length of shapelets, e.g., M. The number of shapelets, i.e., K , of size M, is chosen in a way, such that a large number of input time series, i.e., N, of length Q from different classes C c… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 17 canonical work pages

  1. [13]

    Unsupervised feature learning from time series,

    Q. Zhang, J. Wu, H. Yang, Y . Tian, and C. Zhang, “Unsupervised feature learning from time series,” in Proceedings of the Twenty- Fifth International Joint Conference on Artificial Intelligence , ser. IJCAI’16. AAAI Press, 2016, pp. 2322–2328. [Online]. Available: http://dl.acm.org/citation.cfm?id=3060832.3060946

    work page arXiv 2016

  2. [7]

    k-shape: Efficient and accurate cluster- ing of time series,

    J. Paparrizos and L. Gravano, “k-shape: Efficient and accurate cluster- ing of time series,” SIGMOD Rec., vol. 45, no. 1, pp. 69–76, Jun. 2016. [Online]. Available: http://doi.acm.org/10.1145/2949741.2949758

    work page arXiv 2016

  3. [1]

    Time series shapelets: A new primitive for data mining,

    L. Ye and E. Keogh, “Time series shapelets: A new primitive for data mining,” in Proceedings of the 15th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , ser. KDD ’09. New York, NY , USA: ACM, 2009, pp. 947–956. [Online]. Available: http://doi.acm.org/10.1145/1557019.1557122

    work page arXiv 2009

  4. [2]

    A shapelet dictionary learning algorithm for time series classification,

    J. Zhang, X. Li, L. Gao, L. Wen, and G. Liu, “A shapelet dictionary learning algorithm for time series classification,” in 15th IEEE International Conference on Automation Science and Engineering, CASE 2019, Vancouver, BC, Canada, August 22-26, 2019 . IEEE, 2019, pp. 299–304. [Online]. Available: https://doi.org/10.1109/COASE.2019.8843231

    work page arXiv 2019

  5. [3]

    Clustering of time series subsequences is meaningless: Implications for past and future research,

    E. Keogh and J. Lin, “Clustering of time series subsequences is meaningless: Implications for past and future research,” in In Proc. of the 3rd IEEE International Conference on Data Mining , 2003, pp. 115–122

    work page 2003

  6. [4]

    l2, 1-norm regularized discriminative feature selection for unsupervised learning,

    Y . Yang, H. T. Shen, Z. Ma, Z. Huang, and X. Zhou, “l2, 1-norm regularized discriminative feature selection for unsupervised learning,” in IJCAI, 2011

    work page 2011

  7. [5]

    St-hmp: Unsupervised spatio-temporal feature learning for tactile data,

    M. Madry, L. Bo, D. Kragic, and D. Fox, “St-hmp: Unsupervised spatio-temporal feature learning for tactile data,” 05 2014, pp. 2262– 2269

    work page 2014

  8. [6]

    Learning time-series shapelets,

    J. Grabocka, N. Schilling, M. Wistuba, and L. Schmidt-Thieme, “Learning time-series shapelets,” in Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, ser. KDD ’14. New York, NY , USA: ACM, 2014, pp. 392–

    work page 2014

Show all 17 references
  1. [8]

    Fastdtw: Toward accurate dynamic time warping in linear time and space,

    S. Salvador and P. Chan, “Fastdtw: Toward accurate dynamic time warping in linear time and space,” in KDD workshop on mining temporal and sequential data . Citeseer, 2004

    2004

  2. [9]

    Fast template matching,

    J. Lewis, “Fast template matching,” Vis. Interface, vol. 95, 11 1994

    1994

  3. [10]

    Stochastic neighbor embedding,

    G. E. Hinton and S. T. Roweis, “Stochastic neighbor embedding,” in Advances in Neural Information Processing Systems 15 [Neural Information Processing Systems, NIPS 2002, December 9-14, 2002, Vancouver, British Columbia, Canada] , S. Becker, S. Thrun, and K. Obermayer, Eds. MI...

    2002

  4. [11]

    Visualizing data using t-SNE,

    L. van der Maaten and G. Hinton, “Visualizing data using t-SNE,” Journal of Machine Learning Research , vol. 9, pp. 2579–2605, 2008. [Online]. Available: http://www.jmlr.org/papers/v9/vandermaaten08a.html

    2008

  5. [12]

    Learning a parametric embedding by preserving local structure

    L. van der Maaten, “Learning a parametric embedding by preserving local structure.” Journal of Machine Learning Research - Proceedings Track, vol. 5, pp. 384–391, 01 2009

    2009

  6. [14]

    Self-tuning spectral clustering,

    L. Zelnik-manor and P. Perona, “Self-tuning spectral clustering,” in Advances in Neural Information Processing Systems 17 , L. K. Saul, Y . Weiss, and L. Bottou, Eds. MIT Press, 2005, pp. 1601–1608. [Online]. Available: http://papers.nips.cc/paper/2619-self- tuning-spectral-cl...

    2005

  7. [15]

    Cluster ensembles – a knowledge reuse framework for combining multiple partitions,

    A. Strehl and J. Ghosh, “Cluster ensembles – a knowledge reuse framework for combining multiple partitions,” Journal on Machine Learning Research (JMLR) , vol. 3, pp. 583–617, December 2002. [Online]. Available: http://strehl.com/download/strehl-jmlr02.pdf

    2002

  8. [16]

    Discriminative and uncorre- lated feature selection with constrained spectral analysis in unsuper- vised learning,

    X. Li, H. Zhang, R. Zhang, and F. Nie, “Discriminative and uncorre- lated feature selection with constrained spectral analysis in unsuper- vised learning,” IEEE Transactions on Image Processing , vol. PP, pp. 1–1, 10 2019. Workshop on New Advances in Brain-Inspired Perception,...

    2019

  9. [401]

    Available: http://doi.acm.org/10.1145/2623330.2623613

    [Online]. Available: http://doi.acm.org/10.1145/2623330.2623613

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.