acceptodds
Withdrawn from ICLR 2027

Describe-to-Anchor: Profile-Guided Relational Anchoring for Semantically Grounded Multivariate Time Series Clustering

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deep clustering of unlabeled multivariate time series (MTS) increasingly relies on self-supervised representations derived from reconstruction, augmented views, or evolving pseudo-labels. Without an independent semantic reference, these objectives may capture incidental regularities, drift with the model, or reinforce degenerate geometry. Deterministic temporal descriptors provide stable and interpretable knowledge, but using them as inputs or representations creates a mismatch with fine-grained sequence dynamics. This paper, therefore, proposes PRISM (Profile-guided Representation learning with Intrinsic Structure Matching), a Describe-to-Anchor framework that converts deterministic profiles into fixed relational guidance. Because profile prototypes capture only coarse organization, PRISM complements the prototype anchor with a neighborhood anchor that preserves local distinctions. Since descriptors omit fine-grained dynamics, a temporal patch encoder learns from raw MTS while matching both anchors. This division grounds the learned geometry in explicit temporal characteristics without limiting its representational capacity. Experiments on 30 UEA benchmark datasets demonstrate strong performance against recent baselines, including the best macro-average score and average rank under all four metrics. Statistical tests, anchor-target diagnostics, and semantic traceability analyses on two real-world datasets further validate fixed profile relations.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multivariate time series clustering, representation learning, temporal profiling, relational anchoring, unsupervised learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_koQDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, a Describe-to-Anchor framework for multivariate time series clustering. Deterministic temporal descriptors are converted into fixed relational guidance, while a temporal patch encoder learns from raw MTS and matches both anchors. Although extensive experiments are conducted across 30 UEA benchmark datasets, the technical contribution remains unclear and the comparison to related work is insufficient.

Critical strengths
  1. The experimental evaluation is comprehensive, covering 30 datasets with extensive performance assessment.
Critical weaknesses
  1. The novelty is insufficient and unclear. The lack of a thorough related-work discussion makes the contribution of this work difficult to identify. From the overall framework, the method appears to be a clustering approach that mainly considers contextual information.
  2. Some discussions in the Introduction are not rigorous. For example, the statement that "using profiles as encoder inputs forces compressed summaries and fine-grained dynamics into a shared representation" lacks supporting analysis.
Questions for authors
  1. The related-work discussion is not sufficient, particularly with respect to clustering methods that exploit temporal semantics. The contributions are not clearly positioned.
  2. The proposed strategies, such as prototype learning and neighborhood anchoring, are fairly standard operations that can also be realized by classical approaches, e.g., manifold learning or context-aware clustering. Please clarify the technical differences.
  3. The selected temporal semantics are quite conventional. How can the completeness of these features be guaranteed for representing the underlying multivariate time series?
  4. The proposed framework essentially uses temporal-semantic clustering results to guide data-level clustering. How can the authors ensure that the temporal-semantic types provided are accurate enough and reliable as reference signals?
  5. The paper provides a theoretical analysis of computational complexity, but lacks corresponding experimental results and comparative analysis against other methods. Please supply empirical validation of the claimed complexity.
Reviewer_Bo51OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PRISM, an unsupervised framework for multivariate time series clustering. It extracts statistical profiles from raw sequences and constructs fixed soft assignments to profile prototypes and local neighborhood relations. These serve as supervision targets for training a sequence encoder to reproduce the corresponding relations in the latent space. The final clusters are obtained by applying k-means to the learned representations.

Critical strengths

Clear motivation. The paper motivates fixed profile relations as stable supervision to reduce dependence on evolving pseudo-labels.

Strong performance. PRISM reports the best average ACC, NMI, and ARI across 30 UEA datasets among the compared methods.

Critical weaknesses

Insufficient analysis of the patching strategy. The paper does not analyze sensitivity to the patch size, which may be important. Could the authors compare the current patch setting with variants using different patch sizes on datasets?

Limited failure analysis. Tables 11 and 12 show that Full PRISM performs worse than Profile-only on datasets such as Libras and PEMS-SF, but these performance reversals receive limited discussion. Could the authors analyze these cases and explain why converting profiles into relational supervision leads to worse clustering performance? What dataset characteristics or components of the method contribute to these failures?

Questions for authors

Please see weakness.

Reviewer_4i4BOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PRISM computes statistical profiles of time series and uses their prototype and neighborhood relations as fixed targets to train an encoder on raw signals. It then applies k-means to the learned embeddings to obtain clusters.

Critical strengths

The method intuitively speaking can be more generalizable due to the use of classical human-approved statistical features to provide stable guidance for latent representation learning

Critical weaknesses
  1. The method fundamentally depends on how the profiles are defined, since the chosen statistical features determine the prototype and neighborhood targets. This introduces a strong inductive bias. If these features do not capture the differences that matter for clustering, the model is still trained to follow these fixed relations. And this is the innate bottleneck of this method
Questions for authors
  1. It would be good to compare with a simple baseline that uses the same encoder to directly predict the profile. This would help tell if the gains come from learning these statistical features or from the proposed prototype and neighborhood matching

  2. What direct evidence shows that the learned embeddings retain useful temporal information that is not captured by the profiles?

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

Related papers

Discussion (0)

Sign in to comment.