acceptodds
Under review as a conference paper at ICLR 2027

Relationally Guided History Transfer for Streaming Ensemble Clustering

Abstract

Streaming ensemble clustering seeks to maintain a consensus from base partitions as objects enter and leave successive windows. Existing static ensemble clustering methods, when applied independently to each window, discard historical structure carried by persistent objects. However, directly preserving the preceding consensus can retain outdated structure and provides no historical guidance for newly arrived objects. We propose AGILE, a streaming ensemble clustering framework that uses current base partitions to extend history beyond the window overlap and determine how strongly it should influence the new consensus. Pairwise intersections of base partitions define sparse relations through which historical information reaches arriving objects. To account for redundancy and disagreement among base partitions, AGILE draws complementary evidence from partitions corroborated by their peers and encodes relational corrections with relative-support-weighted signed factors. The corrected current structure supports a history-free consensus against which historical compatibility is assessed. An agreement–coverage gate then regulates the propagated history according to its consistency with the current consensus and cluster representation on the overlap. We establish monotone objective convergence, stationarity of fixed points, and linear time complexity in the window size. Experiments under different window overlap rates demonstrate the effectiveness of AGILE. The source code is available in the supplementary material.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Streaming clustering, Ensemble Clustering, Consensus clustering

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_q7JcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AGILE, which uses current base partitions to propagate historical consensus from persistent objects to newly arrived objects, and regulates how much history influences the new consensus.

The main contributions are a current-first framework, which allows historical information to be extended beyond the overlapping objects, and a relation repair mechanism based on relative-support-weighted signed factors. The repaired relations also provide a current reference without relying on history, which is then used for gating.

Experiments on some datasets, e.g., ten windows, and three overlap rates against fourteen competing methods show that the proposed method achieves SOTA by ACC, ARI, and NMI. The ablation experiments confirming the effectiveness of roles : structural modeling, relation repair, and temporal transfer.

History is treated as a one-step prior that is overwritten by the current consensus, so it cannot accumulate into verifiable, reusable knowledge; what is called "history transfer" is therefore closer to short-term smoothing than to a mechanism that genuinely and continuously guides later windows.

Critical strengths

Strength 1: The current-first strategy prevents stale history from dominating, and it allows a clean fallback to a current-only consensus, yielding the method is more principled than a naive history carry-over.

Strength 2: The framework is complete, and self-contained as well, with convergence, stationarity, and linear-complexity results.

Critical weaknesses

History is only a one-step prior that is overwritten by the current consensus, so earlier structure cannot accumulate and history's influence decays after one step. This makes me unconvinced.

Can the authors summarize, in plain language, what the central finding or conclusion of this work is? The core idea is not intuitive to me.

Questions for authors

Can the authors clarify why one-step history is sufficient in principle, rather than a setting/complexity compromise?

3 comments

Dear Reviewer q7Jc

With the discussion deadline approaching, we wanted to follow up on our rebuttal. We would welcome your thoughts on the clarifications we provided and would be glad to discuss any points you feel need further explanation. Many thanks for the time and care you have put into reviewing our work.

Dear Reviewer q7Jc

With the discussion deadline approaching, we wanted to follow up on our rebuttal. We would welcome your thoughts on the clarifications we provided and would be glad to discuss any points you feel need further explanation. Many thanks for the time and care you have put into reviewing our work.

Dear Reviewer q7Jc

With the discussion deadline approaching, we wanted to follow up on our rebuttal. We would welcome your thoughts on the clarifications we provided and would be glad to discuss any points you feel need further explanation. Many thanks for the time and care you have put into reviewing our work.

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

This paper investigates ensemble clustering for data streams with overlapping windows. Given MM base partitions for each window, the proposed method, AGILE, leverages the persistent identities of objects across adjacent windows and the consensus from the previous window to derive a KK-way consensus for the current window. AGILE first constructs sparse "collision cells" and applies signed relationship refinement to establish a "current-only" reference partition. It then uses the Adjusted Rand Index (ARI) and cluster coverage to gate historical consensus, propagating information to newly arrived objects via the collision cells. Finally, it performs a joint optimization that aligns partitions, corrects relationships, and incorporates temporal constraints. The paper provides theoretical analyses—covering monotonic convergence, fixed-point stability, and linear computational complexity—and reports empirical results (ACC, ARI, and NMI) across multiple datasets and baselines, alongside ablation studies, convergence analysis, and runtime evaluations.

Critical strengths
  1. The problem formulation is clear and practical; it relies solely on base partitions and persistent targets, without requiring historical features or a historical pool of base partitions, making it well-suited for streaming ensemble clustering.

  2. The method is comprehensively designed, featuring a coherent workflow that integrates current-only reference partitions, an agreement-coverage gate, collision-based history transport, and signed relation repair; it also includes closed-form updates and a complexity analysis.

  3. The experimental evaluation is robust, covering multiple datasets, three levels of overlap, and 14 baselines, while also incorporating ablation studies, parameter sensitivity analysis, convergence analysis, and runtime assessment.

Critical weaknesses
  1. Most of the core components are based on existing concepts; the paper primarily focuses on their combination and engineering-level integration. It lacks a clear distinction, both in terms of mechanisms and theoretical foundations, from the most closely related methods in streaming or continual ensemble clustering.

  2. The theoretical guarantees are insufficient to support the paper’s strong claims. The monotonicity established in Theorem 1 relies on fixed relational factors and a "gate" mechanism, while the consensus update employs a surrogate lower bound. The identified fixed point is merely a first-order stationary point; it neither guarantees that incorporating historical information improves clustering quality nor rules out the propagation of errors from the past via "collision cells." Furthermore, because the gate is fixed prior to optimization, it may prematurely determine the influence of historical data.

  3. The experiments are limited to only 10 windows with M=20M=20 and a fixed sampling configuration. Moreover, the paper lacks targeted analysis regarding the design of the gate mechanism, the specific sources of benefit derived from historical data (especially under conditions of low overlap), and the potential for error propagation. Given the current results, the strong conclusion that the proposed method "consistently outperforms 14 other approaches" is not well-substantiated.

  4. The font size in the table has been changed.

Questions for authors
  1. Provide the complete ACC, ARI, and NMI scores, along with their standard deviations, for Table 2, Table 5, and all datasets/overlap rates. Additionally, clarify whether any missing results affect the conclusions drawn.

  2. Where is the direct comparison with the most closely related streaming/continual ensemble clustering methods, using the same inputs, persistent identities, and base partitions? And explicitly state the novel mechanisms of AGILE compared to these existing approaches.

  3. Include an ablation study for the gate (gg): what happens if the coverage term is removed, NMI is used instead, or the gate is updated during the optimization process? In cases of low overlap, do the performance gains stem primarily from the historical data or the current structure?

  4. Does the monotonicity of Theorem 1 rely on the surrogate lower bound and fixed gates/factors? Is the alternating update process (R→w→UR \to w \to U) strictly monotonic in practice, and is it possible that the fixed point reached is merely a poor local optimum?

  5. Quantify the propagation of historical errors: when the prior consensus regarding persistent objects is noisy, or the overlap is extremely sparse, what is the difference, and the statistical significance of that difference, between the variant without the w→Uw \to U update step and the full AGILE framework?

3 comments

Dear Reviewer kUqU

With the discussion deadline approaching, we wanted to follow up on our rebuttal. We would welcome your thoughts on the clarifications we provided and would be glad to discuss any points you feel need further explanation. Many thanks for the time and care you have put into reviewing our work.

Dear Reviewer kUqU

With the discussion deadline approaching, we wanted to follow up on our rebuttal. We would welcome your thoughts on the clarifications we provided and would be glad to discuss any points you feel need further explanation. Many thanks for the time and care you have put into reviewing our work.

Dear Reviewer kUqU

With the discussion deadline approaching, we wanted to follow up on our rebuttal. We would welcome your thoughts on the clarifications we provided and would be glad to discuss any points you feel need further explanation. Many thanks for the time and care you have put into reviewing our work.

Reviewer_hQ4COpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AGILE, a streaming ensemble clustering framework that transfers historical clustering information through sparse relational structures. It combines relation repair, history propagation, and an agreement–coverage gate to balance current evidence and historical consistency. Experiments on eight datasets demonstrate competitive clustering performance.

Critical strengths

The framework is reasonably well designed, integrating current relational structure and adaptive historical transfer within a unified optimization objective, supported by convergence and complexity analyses.

Critical weaknesses

Core components (history propagation, agreement‑coverage gate) lack fine‑grained ablation. Current ablations are limited to few datasets and overlap settings, without comparisons against simple alternative variants. Hence the necessity and practical advantages of these bespoke modules remain unproven.

Questions for authors

To directly verify the effectiveness of history propagation, could the authors separately report clustering metrics for newly arriving objects and persistent objects? Could the authors clarify whether the adopted window model is a mainstream choice for streaming ensemble clustering? Would the proposed method still be applicable under alternative window models or different window‑size configurations?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.