acceptodds
Under review as a conference paper at ICLR 2027

Rondo: Unsupervised Discovery of Recurring Temporal Structure

Abstract

Many real-world time series data exhibit structural properties at multiple scales, from short, recurring units to complex sequences composed of these units. Unsupervised discovery of both these components and structure enables the design of intelligent systems that help interpreting temporal data thereby limiting the amount of costly human annotations required. Existing modeling approaches typically overlook the hierarchical structure inherent to many time series, treating recurring patterns at different temporal scales as independent structures. Moreover, most assume access to the complete data sequence and treat discovery as a static process, limiting their ability to evolve as new observations arrive. We introduce Rondo, an unsupervised approach for modeling recurring hierarchical structure in continuous temporal streams. By explicitly constructing vocabularies of reusable units and their recurring compositions, Rondo captures structure shared across complex temporal patterns while refining and expanding its discoveries as the stream evolves. Evaluations on temporal sequences spanning diverse domains and data modalities show that Rondo outperforms existing unsupervised recurrence-discovery baselines, with particularly pronounced advantages in limited-data and continual-stream settings. These capabilities provide a stronger foundation for recurring-pattern discovery, scalable behavior understanding, and adaptive intelligent systems operating on long, unlabeled temporal streams.

Primary area:
learning on time series and dynamical systems
Keywords:
unsupervised discovery, time-series representation learning, multiscale representation learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_xwo9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces an unsupervised learning approach designed to model and segment the hierarchical and recurrent structure of continuous time series. The pipeline explicitly separates the modelling into two scales:

  1. Local scale: The UnitAlign module uses a self-supervised encoder coupled with a dynamic clustering algorithm (DeepDPM) to learn a vocabulary of basic units.
  2. Global scale: The MotifFormer module relies on a Transformer architecture to learn how these local units assemble to form complex patterns of variable length as texts can do. The method also incorporates a continuous maintenance mechanism that allows the unit and pattern vocabularies to evolve as the time series progress, without having to relearn everything from scratch.

Experiments have been done on 6 different datasets and compared with 6 other methods. Different metric are computed for comparison.

Critical strengths

1- This approach propose to solve the problem of over-segmentation by avoiding the fragmentation of a complex pattern however not grouping in a single pattern a sequence of recurrent units. 2- The method allows for online time series interpretation. 3- The ablation studies shows that both parts of the framework are necessary and articulated. 4- The ablation study on the encoder dependence shows that the methods is overall robust to the representation.

Critical weaknesses

1- The method often fails to reach an interesting SegRation which shows that the methods may tend to segment the signal into smaller units than required. 2- The method depends on many choices (weights in the objective functions, classifier, etc). No ablation study is proposed for them. 3- Several typos are in the References.

Questions for authors

1- The authors mention that in the UnitAligh module, they " alternate between representation adaptation and clustering until the assignments stabilize." Is the convergence ensured? What if the assignment switches form one cluster to another? Is this guiding the creation of new unit? 2- Can you explain the choice of L (eq: 3) and the influence it has on the results? 3- What is the influence of the classifier f? 4- What are the influence of all parameters in the objective functions and the chosen thresholds?

1 comment

Thank you for your patience and constructive feedback. We have completed the additional experiments and analyses outlined in our earlier response and incorporated the findings into the revised manuscript.

  1. Stability. We have clarified UnitAlign’s refinement procedure in Appendix E.2 to make its stopping rule explicit and reproducible. Clustering and adaptor refinement alternate until the ARI between successive partitions reaches 0.995 for two consecutive comparisons, or a maximum of 20 refinement rounds is reached. ARI measures partition consistency independently of arbitrary cluster IDs, allowing the stopping rule to reflect structural changes in the discovered units. We explicitly describe this as an empirical criterion without a theoretical convergence guarantee.
  2. Window length L. Our additional experiments show that RONDO’s comparative gains persist across a broad range of temporal resolutions. We evaluate window lengths of 0.5×, 1×, and 2× the default, together with stride ratios of approximately 1/4, 1/3, and 1/2, across four datasets. Under matched configurations with Time2State and E2USD, RONDO achieves the highest MoF in 17 of 20 configurations and the highest F1@50 in 15 of 20. These results support the robustness of its comparative advantage beyond a narrowly tuned window configuration, while identifying dataset-dependent effects on segmentation granularity. Full results and practical guidance are provided in Appendix A.2.
  3. Mapping f. We have clarified that f comprises the encoder, representation adaptor, and clustering procedure. Table 5 demonstrates that RONDO supports multiple temporal encoders, providing flexibility in the choice of input representation. We further isolate the contribution of representation adaptation through loss-component ablations with encoder representations held fixed (Appendix A.3). Removing pseudo-label-guided supervised contrastive learning consistently reduces MoF, ADR, F1@50, and Edit across all four datasets. This provides direct evidence that UnitAlign’s adaptation contributes to discovery quality beyond the initial encoder representations. Geometry preservation and variance/covariance regularization have more dataset-dependent effects.
  4. Parameter sensitivity. We have added loss-component ablations and sensitivity analyses of the acceptance radius, motif-length range, and window/stride configurations across four datasets (Appendices A.2–A.5). The results support the complementary roles of RONDO’s hierarchical components: pseudo-label-guided contrastive learning strengthens local unit discovery, while view-contrastive learning improves motif representation. RONDO exhibits stable performance across the tested acceptance radii and maintains its comparative advantage over Time2State and E2USD across most window/stride configurations, suggesting that its gains do not depend on narrowly tuned settings. The motif-length analysis further clarifies how the composition range can be adapted to the temporal characteristics of different datasets. Together, these findings support the effectiveness of RONDO’s hierarchical formulation and provide practical guidance for parameter selection.
  5. Over-segmentation. We have directly addressed excessive fragmentation by incorporating a Potts-inspired motif-switching penalty, optimized using dynamic programming, into RONDO’s motif-matching stage. On PAMAP2, this refinement reduces the pre-mapping SegRatio from 18.43 to 3.30 while increasing F1@50 from 18.21 to 71.17, demonstrating that fewer spurious switches are accompanied by substantially improved segment-level quality. Updated results are reported throughout the manuscript, and Table 4 isolates the penalty’s contribution through a variant without regularization. We also report SegRatio before and after semantic mapping and expand Section 4.3 with concrete error analyses. These additions demonstrate RONDO’s improved ability to recover temporal structure while clarifying the dataset-dependent trade-off between suppressing fragmentation and preserving finer transitions.
  6. Related work and references. Thank you for highlighting these issues. We have cross-checked all bibliography entries against Google Scholar exports and corrected several formatting inconsistencies. If there are specific inaccuracies that we may have overlooked, we would appreciate your guidance and will gladly address them.

We hope these additional experiments and clarifications address your concerns and strengthen the evidence for RONDO’s effectiveness and robustness, and we thank you again for your constructive feedback.

Reviewer_9ZK8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces an unsupervised recurring-structure discovery, consisting of UnitAlign for reusable unit discovery and MotifFormer for modeling their higher-level recurring compositions, This paper also demonstrates consistent improvements over 6 segmentation algorithms on the selected four datasets.

Critical strengths

The paper is well presented. It effectively uses figures to present the related work, methodology, and experiments in a logically coherent and concise manner within nine pages.

Critical weaknesses
  1. The idea behind hierarchical temporal modeling appears to be similar to that of the classical segmentation algorithms FLOSS [ICDM 2017] and ClaSS [X]. Both methods directly handle recurring patterns without using deep learning (DL), making them easier to interpret by reducing the number of hyperparameters and likely to offer transferability as well. As online algorithms, these should also be able to handle evolving time series. Furthermore, ClaSS scales linearly with the time-series length N.

[X] Arik Ermshaus et al. 2024. Raising the ClaSS of Streaming Time Series Segmentation. PVLDB 2024

[Y] Arik Ermshaus. 2026. TSSB Data and Baseline Implementations. Supporting website for “Raising the ClaSS of Streaming Time Series Segmentation”. https: //github.com/ermshaua/classification-score-stream

Please compare the performance against ClaSS using the same evaluation settings as [Y]. I would consider it a novel result if the authors demonstrate that using DL improves accuracy. The datasets consist of a small, carefully labeled datasets TSSB(3.5K) and UTSA(12K), and real-world datasets mHealth(34.5k), MIT-BIH-VE(525K), MIT-BIH-VE-Arr(650K), WESAD(2.1M), and SleepDB(3.1M). The numbers in parentheses indicate the median time-series lengths.

The public implementation also appears to be limited to univariate data, but it seems suitable for evaluating the idea behind recurring patterns, as shown in Fig. 1 in your paper.

  1. Several DL-based two-stage methods combining local and global processing already exist in DL-based time-series processing. DC-TSS(DMKD 2025 online) is a segmentation method using AE-based Deep Clustering, while PathTST(ICLR 2023) uses transformer in the second stage, though not for segmentation. For a 2027 paper, please clarify what new insights the proposed DL-based approach offers to readers.
Questions for authors

Please address the weaknesses.

5 comments

Thank you for your patience. We have completed the additional experiments and analyses in response to your comments and incorporated the findings into the revised manuscript.

  1. Task scope and relationship to segmentation. We have revised the Introduction and Related Work to clarify that temporal segmentation is closely related to our task and provides an important evaluation of the recovered structure. RONDO additionally aims to discover which local patterns recur and how they are shared and composed across longer recurring motifs. Its outputs therefore include reusable unit and motif vocabularies alongside their temporal assignments. We have clarified this relationship and expanded the discussion of segmentation, state-discovery, and hierarchical temporal-modeling approaches.

  2. Comparison with segmentation methods. Following your suggestion, we have added comparisons with ClaSS and FLOSS, together with DC-TSS, on TSSB, UTSA, mHealth, and MIT-BIH-VE (Appendix A.1). RONDO leads on mHealth across Covering and change-point F1 and closely approaches ClaSS in Covering on UTSA. However, dedicated segmentation methods perform better on other benchmarks, and the UTSA results show that strong segment overlap does not necessarily imply equally precise boundary localization. These comparisons clarify both the segmentation capabilities and the limitations of RONDO: recurring-composition modeling can recover coherent temporal regions, while precise boundary placement remains an area for improvement.

  3. Insights from explicit hierarchical modeling. We agree that a two-stage architecture alone does not establish novelty. The contribution we seek to demonstrate is the value of explicitly separating reusable local units from their recurring compositions, allowing the two levels to be shared, transferred, and updated separately.

    We now provide complementary evidence for this formulation. The unit-reuse analysis in Section 4.5 identifies local units shared across different activities but organized into distinct recurring compositions. The component ablations show that removing motif-level modeling reduces discovery performance, with particularly substantial segment-level degradation on PAMAP2 and Drive&Act. Finally, the cross-record experiments in Section 4.4 show that the learned unit vocabulary can remain fixed while motif-level discovery adapts to the target recording, further improving MoF and Recall on both evaluated datasets. Together, these results support the practical value of separating transferable local structure from recording-dependent temporal compositions. Thank you again for your constructive feedback. These additions have helped us clarify the contribution and provide a more direct evaluation of its benefits and limitations.

Thank you for your response and revisions. I have a better understanding of the segmentation performance. However, my main concerns remain regarding the novelty and practical value of the proposed method.

  1. Task scope and explicit hierarchical modeling:

ClaSP/ClaSS and FLOSS are nearest neighbor-based methods that internally identify recurring local patterns. Thus, it seems feasible to output these patterns. For example, ClaSP/ClaSS cluster subsequences to identify groups of similar patterns. The Matrix Profile framework underlying FLOSS also supports the extraction of representative subsequence patterns, referred to as dictionaries, and previous studies have investigated their efficient extraction and application to time-series similarity [X, Y].

[X] Yeh, Chin-Chia Michael, et al. "Error-bounded Approximate Time Series Joins using Compact Dictionary Representations of Time Series." SDM, (2022).

[Y] A. Der et al., "Matrix Profile XXVII: A Novel Distance Measure for Comparing Long Time Series," IEEE ICKG (2022).

How do the recurring patterns discovered by Rondo differ from those identified by existing methods, and what practical applications would benefit from these differences?

Identifying recurring patterns seems essentially to be a clustering problem. From this perspective, what is the novelty of Rondo as a clustering method?

  1. Comparison with segmentation methods:
  • Could the authors clarify why SleepDB (3.1M), WESAD (2.1M), and MIT-BIH-Arr (650K) were not included in the evaluation? Given the claim of "scalable behavior understanding" in the abstract, evaluation on larger datasets would be valuable.

  • Would it be possible to include ClaSS and FLOSS in the comparison in Table 1?

Thank you for clarifying these concerns and for pointing us to the related work. We agree that there is substantial overlap between segmentation, state discovery, and recurring-pattern discovery. Although ClaSP/ClaSS and FLOSS primarily target segmentation, they exploit subsequence similarities that can support further pattern extraction, while Matrix Profile dictionaries explicitly provide representative patterns. Indeed, CLaP, which we include as a baseline, uses ClaSP segmentation followed by self-supervised classification and segment-label merging to identify recurring states. Both identifying boundaries before grouping segments and learning temporal representations for clustering are valid approaches to recurring-pattern discovery.

Contribution from a clustering perspective. We agree that grouping or clustering is an important part of recurring-pattern discovery, and its effectiveness depends substantially on the information used to determine similarity. For example, our baselines Time2State and E2USD both learn temporal representations through an encoder and apply clustering to obtain state labels. RONDO does not introduce a new clustering algorithm. Its main contribution is a hierarchical discovery framework that learns complementary representations of local structure and its temporal compositions. Specifically, UnitAlign learns a shared vocabulary of recurring local units through clustering-guided representation adaptation, while MotifFormer models sequences of these units to discover longer recurring motifs. This design allows the same local units to be reused across different motifs, with local structure and its temporal organization modeled at separate but connected levels. RONDO’s contribution lies in this representation and learning mechanism, rather than an exclusive ability to discover recurring patterns.

Practical value. RONDO improves recurring-pattern discovery and reduces fragmentation on several evaluated datasets. Explicitly separating local units from their temporal compositions also supports reuse and adaptation beyond a fixed recording. In continual discovery (Table 2), RONDO shows stronger adaptation to emerging patterns than the evaluated baselines—for example, its final H-score on PAMAP2 exceeds the strongest baseline by 15.46 percentage points. In cross-record experiments, units learned from a source recording can be reused to improve motif discovery in a target recording: updating only MotifFormer on Drive&Act improves MoF from 56.84 to 62.22 compared with direct transfer (Table 3). These results suggest potential benefits for deployments across users or devices that share recurring local structure, where an existing unit vocabulary could support discovery without relearning local representations from scratch. Together, these results support the practical value of maintaining separate but connected unit and motif representations.

Additional benchmark evaluation. We apologize for not covering all suggested datasets. Given the computational resources and time available during the rebuttal period, we prioritized four complementary benchmarks: TSSB, UTSA, mHealth, and MIT-BIH-VE, covering diverse signal characteristics and application domains. We plan to extend the evaluation to the remaining suggested datasets and share further results when available, although completion before the discussion period ends cannot be guaranteed. We appreciate your understanding of these practical constraints.

Comparison in Table 1. We consider ClaSP/ClaSS and FLOSS valuable reference methods. Including them in Table 1 would require an additional grouping stage to convert their segmentation outputs into consistent recurring-pattern labels. Applying their univariate implementations to our multivariate recordings would also require an explicit input adaptation or channel-aggregation strategy. We currently include CLaP as an established ClaSP-based state-discovery baseline, providing a relevant comparison for segmentation-based discovery. We will try to explore a feasible FLOSS-based grouping pipeline and would welcome suggestions on the comparison you would find most informative.

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

This paper introduces RONDO, a framework for unsupervised discovery of recurring hierarchical structure in time series. UnitAlign adapts and clusters local window representations to construct a vocabulary of reusable units, while MotifFormer models their temporal organization to discover higher-level recurring motifs. In the continual discovery setting, the method buffers observations that do not match the existing vocabularies and, once sufficient evidence accumulates, refines existing units and motifs or introduces new ones. The authors evaluate the framework on six temporal datasets, with additional experiments on continual discovery, cross-record transfer, and component ablations. The results show strong frame-level discovery performance, while improvements in temporal segmentation vary across datasets.

Critical strengths

Separating local units from higher-level compositions provides a valuable approach to structural reuse and transfer. RONDO maintains distinct unit and motif vocabularies, allowing local structure to be shared across different higher-level patterns. In Table 3, retaining the learned unit vocabulary while adapting only MotifFormer improves MoF and Recall on both target datasets, supporting the practical value of this design for adaptation without relearning local structure.

Critical weaknesses

The selection of local window configurations is insufficiently justified. UnitAlign uses dataset-specific window lengths and strides, which affect the granularity of the local units used for subsequent motif modeling. The paper does not sufficiently explain how these parameters should be selected without labels, leaving the practical procedure for applying the method to new datasets unclear.

Questions for authors
  1. How are local windows selected without labels? Table 7 specifies different unit window lengths and strides for each dataset. What criteria determine these configurations? Have the authors evaluated how varying window length and stride affects unit discovery and final motif segmentation? This matters for applying the method to a new dataset where activity durations are unknown.

  2. What explains the substantial over-segmentation on some datasets? In Table 1, RONDO obtains SegRatio values of 2.42, 8.04, and 15.37 on MoCap, Drive&Act, and PAMAP2, respectively. Is SegRatio calculated before or after semantic label mapping and merging consecutive intervals with the same mapped label? Given this distinction, do the additional segments primarily reflect meaningful within-activity structure or unstable prediction switches, and which stage of the pipeline contributes most to them?

  3. Could the evidence for unit reuse and motif composition be made more explicit? Appendix A discusses unit persistence, reuse, and activity purity. Could the authors provide the underlying quantitative results and representative examples showing the same unit reused across different motifs? In particular, how do these observations support interpreting the motifs as meaningful compositions of reusable local patterns?

1 comment

Thank you for your patience. We have completed the additional experiments and analyses outlined above and revised the manuscript accordingly.

  1. Window length and stride. We have conducted a sensitivity analysis on four datasets (PAMAP2, OPPORTUNITY, LARa, and Drive&Act), comparing RONDO with Time2State and E2USD under matched window and stride configurations. We evaluate window lengths of 0.5×, 1×, and 2× the default, together with stride ratios of approximately 1/4, 1/3, and 1/2 at the default window length. Each method is evaluated using a single random seed across 20 configurations. Detailed results are provided in Appendix A.2. RONDO achieves the highest MoF in 17 of 20 configurations and the highest F1@50 in 15 of 20. Although absolute performance varies with temporal resolution, particularly for segment-level metrics, RONDO maintains a comparative advantage across most configurations. These results suggest that its improvements are not contingent on a narrowly tuned window configuration. For practical use, we recommend approximately 1–3 seconds of observations as an initial window length, followed by adjustments to the window and stride according to the temporal characteristics of the data.

  2. Over-segmentation. We have refined RONDO’s motif matching stage by introducing a Potts-inspired switching penalty and using dynamic programming to discourage unnecessary motif changes. We have updated the experimental results and included a variant without this penalty in the component ablation study (Table 4) to isolate its contribution. On PAMAP2, for example, the penalty reduces the pre-mapping SegRatio from 18.43 to 3.30, while improving F1@50 from 18.21 to 71.17, demonstrating substantially reduced fragmentation alongside improved segment-level quality.

    The benefits are dataset-dependent, however, and stronger continuity can suppress meaningful short transitions. ActRecTut includes actions shorter than the input window, some of which are missed even without the switching penalty, highlighting temporal resolution as an additional limitation. We therefore assess SegRatio jointly with segmental F1 and Edit, since segment counts alone do not establish boundary or assignment quality. We have expanded Section 4.3 to analyze the remaining over- and under-segmentation errors through concrete examples and discuss their relationship to temporal resolution and regularization.

  3. Unit reuse. We have added a quantitative analysis showing that approximately 37–89% of active units are shared across multiple activity classes (Table 6). We also provide examples illustrating how shared units participate in distinct compositions. In PAMAP2, cycling and rope jumping share a local unit but differ in its repetition and surrounding unit sequence, with both compositions supported by repeated, non-overlapping occurrences. In Drive&Act, magazine and newspaper activities share a unit within different compositions of equal length, showing that their distinction cannot be attributed solely to duration. These observations support the rationale for separating reusable local structure from its higher-level temporal organization. The analysis and examples are provided in Section 4.5.

We hope these additional experiments and analyses address your concerns regarding window sensitivity, segmentation behavior, and unit reuse, while clarifying the strengths and limitations of RONDO. Thank you again for your constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.