acceptodds
Under review as a conference paper at ICLR 2027

Non-redundant Time-series Clustering

Abstract

Traditional time-series clustering methods typically return a single grouping. However, especially with the growing size of time-series data, a single dataset can have more than one meaningful clustering. Therefore, we introduce a non-redundant time-series clustering approach that explicitly uncovers multiple distinct clusterings within the same temporal data. To achieve this, the architecture employs a single-encoder, multiple-decoders design and is trained with a consistency clustering loss. Each decoder is then responsible for a single individual clustering. By independently masking and routing latent representations to these distinct decoders, this design acts as a tool for interpretability. It allows researchers to reconstruct the temporal sequence from isolated embeddings using a dedicated decoder, enabling explicit understanding of the perspectives and structural patterns that drive each separate clustering criterion. Furthermore, we introduce new non-redundant time-series clustering datasets and also collect existing benchmarks that already have multiple labelings. This provides a foundation for future non-redundant time-series analysis. Our evaluation shows that our method outperforms other methods on these datasets.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Non-Redundant Clustering, Alternative Clustering, Time Series Clustering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_4jwnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a non-redundant time-series clustering method that uncovers multiple distinct clusterings in time-series datasets. The proposed method employs a single-encoder and multi-decoder architecture, where each decoder learns to identify distinct clusters using several loss functions, including a non-redundancy loss. Experimental results demonstrate that the proposed method achieves competitive NMI scores across various datasets.

Critical strengths

None

Critical weaknesses
  1. The authors identify two challenging issues in non-redundant time-series clustering, but it is unclear how the proposed method addresses these issues.
  2. It is unclear which component of the proposed method addresses the characteristics of time-series data. In particular, it is unclear how the proposed method differs from existing approaches in addressing time-series data.
  3. Simply extending non-redundant clustering to time-series data offers limited novelty.
  4. The motivation and rationale behind the model architecture are insufficiently explained. The Method section only describes computational processes and training procedures without clarifying why each component is necessary.
  5. It is unclear which components of the proposed method are novel and which are adopted from existing work, making it difficult to assess the novelty of this paper.
  6. The experiments primarily focus on comparing clustering performance using NMI scores. There are limited analyses of the proposed method's distinctive properties and effectiveness.
  7. The paper does not sufficiently explain whether the collected and newly constructed datasets are appropriate benchmarks for evaluating non-redundant clustering performance.
Questions for authors
  1. What limitations does the proposed method aim to address?
  2. Which components of the proposed method are designed to address the characteristics of time-series data? Also, existing non-redundant clustering methods appear to be readily applicable to time-series data. Is there any problem?
  3. What is the rationale behind the model architecture presented in Section 3.1? In particular, why is each component necessary for non-redundant time-series clustering?
  4. Which components are novel, and how do they differ compared to existing methods?
  5. Could the authors provide additional experiments that can directly validate the contribution of this paper?
  6. How did the authors verify that the collected and newly constructed datasets are suitable for evaluating non-redundant clustering? In particular, how did the authors determine the ground-truth number of facets and clusters for each dataset?
  7. In real-world scenarios, the numbers of meaningful facets and clusters are typically unknown. Can the proposed method still discover meaningful patterns when DD and kfk_f differ from their ground-truth values?
3 comments
AuthorsOpenReview ↗

Temporal Design and Novelty, Architectural Evidence, Benchmark Validity, and Count Sensitivity

Thank you for asking us to connect the challenges, design, and evidence. NRTSC discovers and exposes multiple temporal clusterings, with a benchmark for evaluating them.

W1–W2 / Q1–Q2: Challenges and temporal design

Section 1 identifies two difficulties: temporal aspects overlap in the same measurements, and their meanings are unknown in advance. NRTSC learns convolutional temporal features, splits them into facet blocks, and applies sample-specific masks. Each masked block feeds its own clustering head and decoder; the decoded components sum to reconstruct the input (Section 3.1; Figure 2). Reconstruction preserves information; component decorrelation and cross-facet HSIC discourage overlap (Section 3.2). Figure 1 illustrates separate frequency and bump-direction components. Consistency uses perturbed embeddings rather than aspect-specific raw-signal transformations, avoiding the need to prescribe which temporal properties should be invariant. The numbers of facets and clusters are supplied, but their semantic identities are learned. Existing non-redundant methods can process temporal representations: the evaluation explicitly adapts them. ENRC, DivClust, iMClusts, and CMClusts use the same temporal encoder as NRTSC; ISAAC and AutoNR use pretrained temporal-autoencoder embeddings (Section 5; Appendix A.3). This tests the framework beyond encoder replacement.

W3–W5 / Q3–Q4: Novelty and architectural rationale

Sections 2–3 position the contribution in the combination of masked facet routing, separate cluster heads, and dedicated decoders trained through summed reconstruction, consistency, and non-redundancy objectives. The contribution connects each partition to an inspectable temporal component. The components serve distinct roles.

  • Separate blocks and learned masks allow specialization; mask sharpness encourages selective routing and mask usage discourages inactive facets.
  • Consistency stabilizes assignments; balance counters collapse.
  • Squared cosine decorrelation penalizes both aligned and opposing reconstructed components.
  • HSIC compares one facet’s detached predictions with another’s representation, directing its gradient toward representations rather than directly weakening the prediction head (Equations 1–6).
  • Warm-up learns informative components before k-Means initializes the heads for joint optimization (Section 3.3; Appendix A.1).

Such architecture and loss design is novel, additionally none of the previous research has tackled the problem of time non-redundant clustering. Our research introduces the problem, a benchmark for evaluation and a novel design of a method that excel among existing adapted comparison methods.

W6 / Q5: Direct evidence

We included architectural ablation in addition to the loss ablation. The seven-loss ablation spans four datasets and nine labelings (Table 1; Figure 10), with reconstruction, clustering, and balance having the largest effects. Architecture ablation shows the importance of each part. Table 3 reports 8/25 collapsed BumpFreq restarts for the full model, versus 15/25 with fixed routing and 25/25 after jointly removing split, gater, and HSIC. Figure 4 shows that shared decoding can preserve perfect clustering accuracy while mixing signal components, supporting dedicated decoders for faithful interpretation.

Statistical evidence. Figure 5 reports two-sided Wilcoxon signed-rank tests with Holm correction over 45 comparisons per metric (=0.05). NRTSC significantly outperforms DivClust, ISAAC, iMClusts, CMClusts, and DTCC in both NMI and ARI, and AutoNR in ARI.

W7 / Q6: Benchmark suitability and reference counts

Section 4 and Appendix A.2 distinguish six real-data-derived datasets, two existing synthetic datasets, and three constructed datasets. Exact signal matching reunites documented annotations for the same real observations. SyntheticControl and TwoPatterns factor documented class definitions without altering the signals. BumpFreq and BME-AC supply controlled factors; Plane-AC adds a controlled trend to aircraft outlines. The reference facet count is the number of available criteria and each cluster count is its label cardinality (Tables 2 and 5).

Q7: Unknown or misspecified counts

NRTSC does not estimate counts automatically. Appendix A.5/Table 8 tests overestimates on four datasets over three seeds. DistalPhalanx age/TW NMI changes only from 0.415/0.620 at [3,6] to 0.393/0.601 at [4,7] and 0.413/0.617 with an extra binary facet. BME-AC retains perfect recovery with an extra facet, but three clusters per facet make peak recovery unstable. SyntheticControl direction remains at NMI 0.685–1.000 and GunPoint era at least 0.781 across tested settings. Extra facets can duplicate partitions, and some clusters remain unoccupied. Therefore, meaningful structure can survive overestimates.

Thank you for the detailed rebuttal and additional experiments. However, I still have several concerns regarding the contributions of this paper.

It remains unclear what makes the proposed method specifically designed for non-redundant time-series clustering. Beyond the temporal encoder, the authors do not clearly identify which components are proposed to address challenges specific to time-series data. I also have a question about whether the components described in "W3-W5 / Q3-Q4" address challenges specific to time-series data.

If there are no components specifically designed for time series, comparisons on other modalities, such as images, should be conducted to validate the contributions of the proposed method.

Although the authors claim novelty in the proposed architecture and loss design, the specific novelty of the loss functions remains unclear. In particular, how do the three losses differ from those used in existing non-redundant clustering methods?

The additional experiments raise further concerns about the reliability of the proposed method. The authors reported collapse in 8 out of 25 runs, even with the full model. Such a high failure raises questions about the stability and reliability of the proposed method.

Considering these unresolved concerns, I will maintain my original score.

The design choices are based on the temporal structure of the data and we explain a few of them below, however The relevant question is whether their combination provides benefits over existing methods adapted to the same setting. Our comparisons use the same temporal encoder for the deep non-redundant baselines, and NRTSC achieves the best NMI and ARI ranks while providing inspectable temporal reconstructions. These results support the design’s effectiveness for non-redundant time-series clustering.

Design choices relevant to temporal structure of the data:

  • Additive reconstruction: each facet reconstructs a full-length component, and their sum (superposition) reconstructs the input.
  • Each facet’s representation feeds both a clustering head and a dedicated decoder. This connects its partition to an inspectable waveform, revealing the temporal pattern associated with that clustering.
  • Decorrelation of reconstructed waveforms: Squared cosine similarity discourages facets from reconstructing aligned components or opposing components that cancel in the sum. This supports complementary explanations of the observed signal

How our design choices differ from existing methods?

  • ENRC learns a single parameter matrix whose softmax produces feature weights shared across samples. Our gater extends this weighting idea to sample-dependent routing: an MLP predicts masks from each input’s embedding and applies them to separate facet blocks before clustering and decoding. This is a conceptual extension, not an identical parameterization: ENRC weights shared transformed coordinates, whereas NRTSC weights coordinates within facet-specific blocks.

  • ENRC uses reconstruction through a shared decoder. iMClusts also combines autoencoding with diversified attention to learn salient clustering subspaces; consequently, neither reconstruction nor attention-based feature selection alone constitutes our novelty. In NRTSC, each facet representation feeds its own decoder, and the sum of the decoded components reconstructs the input. Mask-sharpness and mask-usage penalties encourage selective routing and discourage inactive masks. This couples representation learning to an inspectable temporal decomposition rather than only to the recovery of the complete input.

  • Clustering: prediction consistency rather than centroid compression. Methods such as ENRC, AutoNR, and CMClusts use centroid-based objectives or k-Means-related procedures that favor compact clusters in their respective representation spaces. NRTSC instead learns assignments through consistency between predictions from perturbed latent views, together with a balance term that discourages collapse. Consequently, it does not explicitly require embeddings to concentrate around cluster centroids.

The final clustering does not collapse.

Here, “collapse” means that facets produce nearly identical partitions (inter-facet NMI > 0.95), rather than all samples being assigned to one cluster. The 8/25 figure counts individual restarts before model selection, across five seeds with five restarts each. It therefore measures redundancy within individual restarts, not the failure rate of the complete procedure. Table 7 in Appendix A.3 reports the average inter-facet NMI between the clusterings discovered by each algorithm including NRTSC. For most datasets the inter-facet NMI is near zero, and the maximum inter-facet NMI is only 0.39.

NRTSC selects among restarts using an unsupervised score. Table 3 reports selected-model recovery alongside collapse counts across all restarts, making initialization sensitivity visible.

The controlled comparison also shows that the proposed design improves stability: collapsed restarts increase from 8/25 for the full model to 15/25 with fixed routing, 16/25 with a shared decoder without the gater, and 25/25 when the split, gater, and HSIC are jointly removed. Thus, learned routing and facet separation substantially reduce redundant solutions, although they do not eliminate them.

We acknowledge the remaining initialization sensitivity as a limitation. Our results support effectiveness under the reported restart-and-selection protocol, rather than a guarantee of successful non-redundant recovery from every initialization.

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

The paper proposes NRTSC for non-redundant clustering of univariate time series, aiming to discover several meaningful partitions without prescribing what each partition should represent. NRTSC applies learned masks to separate latent blocks produced by a shared encoder and assigns a clustering head and a decoder to each facet. The decoded components are summed to reconstruct the input while consistency and balancing losses work together with non-redundancy regularization to encourage useful and distinct partitions. The number of facets and the cluster count for each facet are provided. The authors also assemble a benchmark comprising 11 datasets with a total of 25 reference labelings. Comparisons with nine baselines show the best reported average NMI and ARI ranks. The main contribution is a time-series formulation that links multiple discovered clustering criteria to inspectable reconstructions, supported by a reusable benchmark and empirical evaluation.

Critical strengths

Discovering multiple interpretable clustering views.

Conventional time-series clustering typically returns a single partition and may overlook other meaningful structure within the same data. NRTSC addresses this limitation by learning multiple clustering views without predefined semantic criteria. Learned masks and separate clustering heads allow different latent representations to support different partitions. Each view also has a dedicated decoder that makes the encoded temporal patterns inspectable. The reconstructed components are combined to reconstruct the input while non-redundancy objectives discourage overlapping information across views. This design supports both the discovery of alternative groupings and the inspection of the temporal structure associated with each grouping.

Best average NMI and ARI ranks across the benchmark.

The evaluation includes 11 datasets with 25 reference labelings and comparisons against 9 baselines. NRTSC achieves the best average NMI rank of 2.66 and ARI rank of 3.06. ENRC ranks second with 3.24 and 3.80 respectively. The improvements extend beyond the newly constructed datasets. NRTSC obtains higher mean NMI and ARI scores than ENRC on all six reference labelings from the three Phalanx datasets. The main comparison also reports results over five runs and uses the same temporal encoder for the deep non-redundant baselines. These results and experimental choices provide convincing support for the method’s overall performance advantage.

Critical weaknesses

Recovery of individual clustering criteria is uneven.

The strength of the results varies across criteria. NRTSC obtains NMI 0.09 for shape and 1.00 for direction on SyntheticControl. And it trails ENRC on all three reference partitions on GunPoint. The reported advantage is therefore an aggregate improvement with effectiveness varying across datasets and criteria. These results suggest caution in generalizing the method's ability to recover several meaningful factors simultaneously. A discussion grounded in the existing results would help clarify the scope of the gains. The overall comparison remains positive despite these limitations.

Meaningful clustering criteria may retain shared information.

The interpretation of diversity would benefit from a more nuanced discussion when useful criteria are related. Figure 6 includes dependent or nested reference labelings, such as day-of-week and weekend, or subject and gender, while the regularization and restart selection favor reduced dependence. Lower inter-facet NMI alone does not establish that the resulting partitions are more meaningful. Avoiding duplicate partitions should be distinguished from removing legitimate shared structure. The authors already acknowledge that linear HSIC does not guarantee full independence. Discussing this additional interpretive trade-off would help readers assess Table 5 and the recovered facets.

Questions for authors
  1. How can the difference between the two SyntheticControl criteria and the performance relative to ENRC on GunPoint be explained?
  2. Based on the current results and the method's design, what kinds of temporal structure appear most suitable for NRTSC?
  3. Figure 6 shows that some reference labelings are correlated, such as day of the week and weekend status. If the model correctly recovers both labelings, its output partitions will naturally be correlated. Is lower inter-facet NMI necessarily better in this case? Please clarify how these meaningful correlations should be considered when interpreting Table 5 and using the diversity term to select the final model from multiple training restarts.
1 comment
AuthorsOpenReview ↗

Explaining Uneven Recovery, Suitable Temporal Structure, and Meaningful Facet Dependence

We thank the reviewer for the thorough and helpful feedback.

W1 / Q1: Uneven recovery on SyntheticControl and GunPoint

On SyntheticControl, Table 2 reports NMI 0.09 for shape and 1.00 for direction. Appendix A.6 examines this asymmetry: one facet consistently recovers rising versus falling direction (ARI 1.00), whereas the other varies across runs, with average pairwise ARI 0.19 (0.17 for ENRC). The shape distinction is local, its sign depends on direction, and the step may occur between t = 20 and t = 40. It explains only 1.9% of signal variance, compared with 65% for direction. These observations help explain why the dominant criterion is recovered reliably while the reference shape partition is not. The decoders also expose what the less stable facet captures. In one run, it detrends gradual series, while its outputs for sudden shifts depend on timing and direction. Its partition agrees with the XOR of the reference labels at ARI 0.92 (Section 5; Appendix A.6). This is a post hoc interpretation, not an improved reference-label score or evidence that the shape criterion was reliably recovered. It illustrates the value of inspecting the discovered partition rather than assuming its meaning from the requested facet count.

On GunPoint, ENRC indeed outperforms NRTSC on all three reference partitions in Table 2. Appendix A.6 shows that NRTSC primarily follows absolute hand position: across five runs, all 15 facets can be closely described by thresholds on resting, peak, or mean position, with cross-validated agreement ARI 0.81 ± 0.12. Recording era is recovered strongly in four runs (ARI 0.81–0.99), but only at 0.40 in the fifth; the other facets vary across runs (pairwise ARI 0.28). Thus, our analysis explains what the model learns and why era is easier, while acknowledging that interpretable position-based partitions do not establish recovery of action or gender. This limits any claim of uniformly recovering several semantic factors.

Q2: Suitable temporal structure

The exploratory analysis in Appendix A.6 associates reference-criterion recovery with its share of variance at fixed time points (Spearman ρ = 0.87). Most criteria explaining under 5%, including event order, timing differences, and variable-time steps, remain at ARI 0.12 or below. The present evidence therefore favors salient structure aligned in the observed signal. It does not establish a causal mechanism or a general recovery rule, since criteria within a dataset are related. This qualification is explicit in the appendix.

W2 / Q3: Meaningful dependence versus duplication

We agree that lower inter-facet NMI is not automatically better when reference criteria are related. Day-of-week and weekend, or subject and gender, should retain dependence when correctly recovered (Appendix A.2; Figure 7). Section 5 explicitly requires inter-facet NMI to be interpreted alongside recovery scores. Table 7 measures dependence between predictions; by itself it cannot establish their semantic quality.

However, the model uses soft regularization rather than a hard independence constraint. Sample-specific masks weight separately learned latent blocks, while reconstruction, consistency, and balance preserve signal information and encourage useful, non-collapsed assignments (Sections 3.1–3.2). Linear-kernel HSIC penalizes cross-covariance and does not guarantee full statistical independence; component decorrelation provides a complementary signal-domain penalty (Equations 4–5). Accordingly, the learned outputs can retain dependence: Table 7 reports NRTSC inter-facet NMI of 0.318 on DodgerLoop and 0.348 on SemgHandCh2.

Restart selection likewise combines mean silhouette quality with mean diversity, 1 − NMI, with default diversity weight 0.5; it does not minimize dependence alone, and reference labels are not used for selection (Appendix A.1). Related criteria remain an interpretive trade-off: the diversity score must be read together with reference recovery and decoded components. The defensible conclusion is that NRTSC can discover and expose complementary temporal partitions, with strong aggregate performance, while dominant signals, related criteria, and run-to-run variability constrain recovery of particular reference factors.

Reviewer_bEv2OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework to deal with non-redundant time series clustering. Specifically, it uses a shared encoder to decompose the given time series into different facets with different learning heads. Working with designed loss function, the whole framework is expected to extract implicit temporal patterns from one given time series into several branches. Experiments show its effectiveness.

Critical strengths
  1. Overall, the draft is easy to follow and it is clear to understand.
  2. Non-redundant time series clustering is a critical and practical research topic for real world application.
  3. Using a shared encoder with different heads to decompose given time series is a reasonable design
  4. Experimental results support its statement.
Critical weaknesses
  1. Comparison methods are relatively old. The author may want to add more recently published papers to compare with, which further strength the paper claim.
  2. Performance gains are relatively small, then it is hard to strongly support the proposed method works better than existing methods.
  3. More ablation study is need to detailedly decompose the proposed framework and show each part is really working.
  4. The author may want to revise the table and figures parts of the paper to make them dense and more informative.
Questions for authors

Please check the section above.

1 comment
AuthorsOpenReview ↗

Recent Baselines, Performance Gains, Component Ablations, and Improved Figures and Tables

We thank the reviewer for the constructive suggestions. The paper supports NRTSC through comparisons with recent methods, loss and architecture ablations, and analyses of the temporal components associated with each clustering. The newly added parts are highlighted in yellow.

W1: Recent comparison methods

The evaluation includes CMClusts (2025), DivClust (2023), iMClusts and AutoNR (2022), as well as DTCC (2023), TS2Vec + k-Means (2022), and k-Shape (Section 5). To test more than the benefit of a temporal encoder, ENRC, DivClust, iMClusts, and CMClusts use the same encoder as NRTSC; ISAAC and AutoNR receive embeddings from a pretrained temporal autoencoder (Appendix A.3). Sections 2.2 and 5 also explain the distinction from recent guided methods such as Multi-MaP, Multi-Sub, and MLLMMC: these require user-specified semantic criteria or language-model-generated descriptions. NRTSC receives the numbers of facets and clusters, but not their semantic meanings. Such additional guidance would change the comparison setting.

W2: Scope of the performance gains

Our claim is an aggregate advantage across 11 datasets and 25 reference labelings, rather than superiority on every criterion. NRTSC has the best average ranks for both NMI and ARI: 2.66 and 3.06, versus 3.24 and 3.80 for ENRC, the next-best method (Figure 5; Tables 2 and 6). It ranks first on 13/25 labelings when ties count for every tied method (Appendix A.3). The exploratory Wilcoxon–Holm comparisons support improvements over DivClust, ISAAC, iMClusts, CMClusts, and DTCC in both metrics, and AutoNR in ARI. Figure 5 explicitly notes that related labelings make these tests exploratory. The evidence therefore supports strong overall performance, with the limits of the comparison stated explicitly.

W3: Loss and architecture ablations

The current version of the paper has two different ablation experiments in the experiments section. Table 1, and Appendix A.4/Figure 10 ablate all seven loss terms across four datasets and nine reference labelings over three runs. Removing reconstruction reduces NMI on every labeling. Removing clustering reduces BME-AC plateau and BumpFreq frequency NMI by 0.97 and 0.61; removing cluster balance reduces Plane-AC aircraft NMI by 0.65. The remaining regularizers have smaller, criterion-dependent effects, so we do not claim that every term independently produces a large accuracy gain.

The architecture study additionally compares four variants over five seeds and five restarts (Table 3). On BumpFreq, the full model has 8/25 collapsed restarts, compared with 15/25 for fixed routing and 25/25 when the split, gater, and HSIC are removed together; collapse means inter-facet NMI > 0.95 before restart selection. The latter variant collapses despite achieving the lowest reconstruction error on three datasets. This directly demonstrates why accurate reconstruction alone is insufficient for distinct clusterings and supports the combined role of facet separation and learned routing. Dedicated decoders serve a different purpose: faithful inspection of the clustering criterion. Figure 4 shows a BumpFreq example in which both full and shared-decoder models attain NMI 1.00 for both labelings and almost identical total reconstruction error, yet the shared decoder mixes the components. For the illustrated series, bump/sinusoid NMSE increases from 0.013/0.020 to 0.206/0.375. The result distinguishes clustering accuracy from the faithfulness of its reconstructed explanation.

W4: Clearer tables and figures

We have updated figures 1 and 2, and added two new figures to the main papers. In addition to updating other figures in the appendix. Similarly, we worked on the tables to make them denser and clear to deliver the important information directly. Figure 2 displays the shared encoder, masked facet pathways, clustering heads, separate decoders, and summed reconstruction; Figure 6 details the training phases and losses. Table 1 summarizes the main loss effects, Table 3 reports architecture results and collapse counts, and Figure 4 makes component leakage visible. Together with the full loss heatmap in Figure 10, these provide compact evidence for how the design works and where its benefits arise. Beyond numerical gains, the paper contributes an eleven-dataset benchmark and a way to inspect the temporal patterns behind multiple clusterings (Sections 1 and 4).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.