Pith. sign in

REVIEW 5 major objections 4 minor 56 references

Federated clustering's usual Non-IID assumption misses clusters that are fragmented across clients; this paper's GOLD framework learns fine-grained local clusters and merges them globally in a single round, outperforming eight existing meth

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 08:15 UTC pith:LTKB4CQF

load-bearing objection Fragmented-cluster problem framing is fresh, but the final clustering stage is undefined, so the empirical claims don't yet stand. the 5 major comments →

arxiv 2601.17512 v1 pith:LTKB4CQF submitted 2026-01-24 cs.LG

One-Shot Federated Clustering of Non-Independent Completely Distributed Data

Yiqun Zhang , Shenghong Cai , Zihua Yang , Sen Feng , Yuzhu Ji , Haijun Zhang This is my paper
classification cs.LG
keywords federated clusteringNon-IIDNon-ICDone-shot federated learningcompetitive penalized learningmulti-granular clusteringautomatic cluster numberedge computing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that the standard Non-IID assumption in federated clustering misses a more realistic failure mode: a single global cluster can be fragmented across clients at different granularities, and a client can hold non-adjacent subclusters of the same cluster. To handle this 'Non-Independent Completely Distributed' (Non-ICD) condition, it introduces GOLD, a one-shot federated clustering framework. GOLD has clients find compact micro-clusters locally (FCPL), the server recursively merges these into multi-granular global distributions (MCPL), and finally encodes the multi-granular assignments into an enhanced representation for the target clustering (REMC). The paper claims GOLD outperforms eight existing FC methods across ten datasets, remains robust as the Non-ICD degree grows and as the number of clients scales, and produces interpretable nested cluster relationships without knowing the number of clusters in advance.

Core claim

On the paper's own terms: real-world federated data commonly violates the implicit completeness assumption behind existing federated clustering methods. The paper defines Non-ICD as the condition where a global cluster is decomposed into subclusters at different granularities that are dispersed across clients (granularity misalignment and non-adjacent subclusters). Under this condition, the paper's GOLD framework claims to achieve accurate global clustering with only a single round of communication by applying competitive penalized learning twice: fine-grained CPL at each client discovers compact micro-clusters without a pre-specified k, and multi-granular CPL at the server recursively aggre

What carries the argument

Competitive Penalized Learning (CPL): candidate clusters compete for data objects; winning clusters are rewarded and the nearest rival is penalized, so redundant clusters are eliminated and the number of clusters emerges automatically. GOLD uses CPL at two levels: FCPL (fine-grained CPL) on clients to extract compact micro-clusters from fragmented local data, and MCPL (multi-granular CPL) on the server, which recursively re-runs CPL while inheriting only the previous cluster count, producing a set of clusterings at different granularities. A feature-cluster importance matrix H weights features per cluster, and the final stage, REMC (Representation Enhancement Based on Multi-Granular Clusters

Load-bearing premise

The final REMC stage treats cluster memberships encoded as integer codes and measures similarity with bitwise operations (XOR and negation), but the paper never defines these operations for the real-valued centroids that the alternating optimization produces, and the encoding is not invariant to arbitrary relabeling of clusters; if this similarity is ill-defined or label-dependent, the reported gains do not come from a fully specified algorithm.

What would settle it

Reproduce REMC as written on a small dataset: run FCPL and MCPL to obtain the Q matrices, then attempt to compute Eq. (25) using the centroid of a cluster in X^(e). Since X^(e) entries are integers but centroids c_j are averages (real numbers), the XOR/negation operations are undefined; a reader should be able to point to the line in the paper where these operators are defined for real inputs. Alternatively, permute the cluster indices of a fixed clustering before encoding: if GOLD's final partition or its objective value changes, the method is not relabeling-invariant.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A single communication round suffices: clients upload only local centroid sets, reducing communication from O(r·c) to O(c) and shrinking the privacy attack surface.
  • The number of clusters need not be provided: CPL's competition mechanism prunes redundant candidates, so k* is discovered rather than assumed.
  • The learned multi-granular structure yields interpretable nested cluster relationships, revealing how coarse global clusters decompose into fine-grained local subclusters.
  • Performance degrades more gently than baselines as the Non-ICD degree increases, and remains stable when the federation grows from 100 to 1000 clients.
  • The design inherits convergence from CPL and from weighted k-means, ensuring alternating updates terminate; complexity stays linear in n, d, and k0.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If Non-ICD is as pervasive as the paper argues, the multi-granular output of GOLD could double as a hierarchical clustering view of the global data, useful for downstream federated tasks such as personalized federated learning that needs alignment between local subclusters and global clusters.
  • The λ measure, defined as mean pairwise Jensen-Shannon distance between client density estimates, could be adopted as a diagnostic statistic for data heterogeneity in federated systems; a testable extension is to verify whether λ correlates with performance drops of existing methods on real deployment logs.
  • The REMC encoding in Eqs. (23)-(28) implicitly assumes cluster labels are comparable across granularity levels; since label indices are arbitrary, an editor-level extension would be to replace the bitwise similarity with a permutation-invariant distance (e.g., contingency-based similarity) and test whether performance improves or remains equal.
  • The simulation creates Non-ICD by sub-partitioning clusters via k-means; a natural extension is to evaluate on naturally fragmented data, such as sensor streams where each device sees only one room, to validate whether the benefits persist beyond the synthetic protocol.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper introduces Non-ICD (Non-Independent Completely Distributed) as a generalization of Non-IID for federated clustering, and proposes GOLD, a one-shot framework with three phases: FCPL (Fine-grained Competitive Penalized Learning) on clients, MCPL (Multi-granular Competitive Penalized Learning) on the server, and REMC (Representation Enhancement based on Multi-Granular Clusters) for final clustering. GOLD claims to automatically determine cluster numbers at multiple granularities, to aggregate fragmented client clusters, and to outperform eight baselines across ten datasets on multiple clustering indices. The paper also proposes a measure lambda for the degree of Non-ICD.

Significance. If the result held, the paper would address a real and under-explored problem in federated clustering: the fragmentation of global clusters across clients at different granularities. The one-shot communication design and the interpretable multi-granular output would be practically valuable for IoT applications. The empirical study is broad, covering ten datasets, eight baselines, significance tests, ablations, and scalability experiments. However, the central final clustering stage (REMC) is not well-defined, the claimed automatic k* selection is contradicted by the experimental protocol, and the complexity analysis mis-handles k0. These are load-bearing issues, not presentation concerns. The method builds on the external CPL framework and weighted k-means rather than fitting target labels, so there is no circularity, but the paper is not currently verifiable as a clustering algorithm.

major comments (5)
  1. [III-D, Eqs. (23)-(28)] The REMC stage is not a well-defined clustering objective. Eq. (23) encodes cluster membership as the integer label x_{iδ}^{(e)} = Σ_j j q_{ij}; permuting the columns of Qδ changes every encoded feature and therefore the final objective. Eq. (25) defines s(x_i^{(e)}, C_j) = ||u_j · ¬(x_i^{(e)} ⊕ c_j)||_2 but ⊕ and ¬ are never defined for integer codes and (possibly real-valued) centroids c_j. No update equation for c_j appears in Algorithm A.2; the centers remain at their random initialization. Eq. (28) uses ¬(x⊕x') as an equality test, but bitwise XOR on arbitrary integer labels induces a Hamming metric that depends on the binary representation rather than cluster identity. Consequently the final partition is not invariant to cluster relabeling and the objective is not a function of the data. This stage produces GOLD's reported clusters, so the experimental gains cannot be attributed to
  2. [IV.A, Table I] The paper claims automatic determination of k*, but Section IV.A states 'For all compared methods, the global number of clusters k* searched at the server is set according to the benchmark labels of the datasets reported in Table I.' The REMC objective in Eq. (24) and Algorithm A.2 (Phase 3) require the target k* as input (U ∈ R^{Δ×k*}, assignments over 1..k*). Thus the reported experiments evaluate GOLD with the true k*, not an automatic procedure. The automatic-k* claim needs either a protocol in which k* is not supplied, or a precise statement of which stage is automatic and which uses the benchmark k*.
  3. [III.E, Theorem 1; Section IV.A] The complexity analysis treats k0 as constant, but Section IV.A sets k0 = 0.5n(l) for clients and k0 = 0.5n for the server. Lemma 2 discards k* and Δ as 'relatively small' but k0 is linear in n; under the experimental setting the server-side cost is O(d n k0) = O(d n^2) and the client cost is O(d L n(l) k0) = O(d L n n(l)). Fig. 7 therefore does not verify the claimed linear complexity unless k0 is held fixed independently of n. The theorem and the scalability experiments must be revised to reflect the actual scaling of k0.
  4. [III.B, Eqs. (8)-(9)] Eqs. (7)/(8) define s as the (weighted) Euclidean distance ||x_i - c_j||, yet Eq. (9) selects the winning cluster by v = arg max_j γ_j w_j s(x_i,C_j). Maximizing a distance assigns objects to the farthest cluster, which is contrary to the clustering objective in Eq. (6) and to the later use of exp(-0.5·distance) in Eq. (22) as a similarity. This must be either s := -distance (or exp(-distance)) with arg max, or arg min over distances. As written, FCPL is not a valid clustering update.
  5. [III.C, Algorithms 1/A.2] MCPL's recursion is not specified in the pseudo-code. Algorithm A.2 line 9 calls 'Run Algorithm A.1 with cluster centroids matrix X', but Algorithm A.1 initializes k0 random centroids and receives no inherited k. The termination condition in Algorithm 1 references k_old and P_old, but neither is initialized before the loop, and P_old is never updated. The multi-granular Q produced by MCPL is therefore not reproducible from the algorithms as written.
minor comments (4)
  1. [III.C, Eq. (22)] The denominator sums over t with h_j in the exponential for each t, rather than h_t. Please clarify whether this is intentional or a typo; as written, the normalization is not a standard softmax over clusters.
  2. [III.A, Eqs. (2)-(4)] \hat f is a KDE density, not a probability mass on the observed sample set, so Eq. (4) is not literally a KL divergence over the data points. State the normalization or discretization used to make lambda a valid JS distance.
  3. [Appendix II.D] The hyper-parameter evaluation omits k0 = 0.1n and 0.2n. This should be stated in the main text, or results provided; otherwise sensitivity at small initial cluster counts is untested.
  4. [Fig. 7(b)] The x-axis label reads 'No. of Object d' but appears to mean 'No. of Features d'. Please correct.

Circularity Check

0 steps flagged

No circularity: GOLD's derivation chain rests on external CPL and weighted-k-means; self-citations are contextual, and REMC's underspecified operators are a correctness gap rather than a circular reduction.

full rationale

Walking the derivation chain: Definition 1 defines Non-ICD as a data condition, and λ (Eqs. 1-4) is an externally defined JS/KDE quantification; neither encodes the output labels. FCPL and MCPL are built on CPL [44], an external algorithm, with feature importance from standard Hellinger distances; the convergence argument imports [44] and weighted-k-means [51], both external. No equation fits the final partition and then reports it as a prediction. REMC (Eqs. 23-29) encodes MCPL's affiliation matrices into X(e) and clusters that code; the result is a post-processing ensemble step rather than an identity, so it does not reduce to its input by construction. The many self-citations [34-40] appear in related work or as baselines and are not load-bearing for GOLD. The experimental protocol gives the true k* to all methods, which weakens the unknown-k* claim, but this is an evaluation-design issue, not circularity. The REMC stage is indeed underspecified—bitwise ⊕/¬ is never defined for real centroids and no centroid update equation is given—so the final objective is not fully specified; however, that is a correctness/reproducibility problem, not a demonstration that the derivation is equivalent to its inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical or latent entities are postulated; Non-ICD is a distribution-level concept, and clusterlet centroids are just centroids. The free parameters are initialization/stopping choices and a bandwidth for the auxiliary λ measure.

free parameters (4)
  • initial cluster count k0 = 0.5 n(l) (client), 0.5 n (server)
    Chosen as a heuristic multiple of dataset size; used for FCPL and MCPL initialization; hyperparameter sweep in Appendix II.D shows a plateau but no principled selection.
  • learning rate η = 0.05
    Set based on Appendix II.D evaluation on the same benchmark datasets; controls CPL weight updates.
  • MCPL stopping threshold ε = not specified ('small threshold')
    Stopping condition for recursive granularity exploration; value not reported, so convergence criterion is not reproducible.
  • KDE bandwidth h = cross-validated likelihood
    Used to estimate client densities f^(X(l)) for λ; not used in GOLD itself, but λ is the Non-ICD measure in Table I and simulation.
axioms (4)
  • domain assumption CPL (Rival Penalized EM) converges and eliminates redundant clusters as established in [44]
    FCPL and MCPL convergence is asserted by inheritance from CPL [44] (Section III.E); the objective modifications (feature-cluster weights, normalized similarity) are not proven to preserve the monotonicity.
  • ad hoc to paper Cluster-ID integers encoded in Eq. (23) provide a meaningful similarity space
    REMC assumes relabeling cluster indices is not arbitrary and that Euclidean/bitwise distances on these codes are informative; no invariance or justification is given.
  • domain assumption Real-world clusters decompose into non-overlapping subclusters that can be simulated by k-means partitioning
    The Non-ICD simulation (Section IV.A, Steps 1-5) uses k-means to create subclusters; the evaluation only tests this operationalization of Non-ICD.
  • domain assumption Uploading centroids is sufficient privacy protection
    GOLD's privacy claim rests on sending only centroids; standard FL caveat that centroids can leak information (e.g., membership inference) is not addressed quantitatively.

pith-pipeline@v1.3.0-alltime-deepseek · 36749 in / 12743 out tokens · 134679 ms · 2026-08-03T08:15:07.481381+00:00 · methodology

0 comments
read the original abstract

Federated Learning (FL) that extracts data knowledge while protecting the privacy of multiple clients has achieved remarkable results in distributed privacy-preserving IoT systems, including smart traffic flow monitoring, smart grid load balancing, and so on. Since most data collected from edge devices are unlabeled, unsupervised Federated Clustering (FC) is becoming increasingly popular for exploring pattern knowledge from complex distributed data. However, due to the lack of label guidance, the common Non-Independent and Identically Distributed (Non-IID) issue of clients have greatly challenged FC by posing the following problems: How to fuse pattern knowledge (i.e., cluster distribution) from Non-IID clients; How are the cluster distributions among clients related; and How does this relationship connect with the global knowledge fusion? In this paper, a more tricky but overlooked phenomenon in Non-IID is revealed, which bottlenecks the clustering performance of the existing FC approaches. That is, different clients could fragment a cluster, and accordingly, a more generalized Non-IID concept, i.e., Non-ICD (Non-Independent Completely Distributed), is derived. To tackle the above FC challenges, a new framework named GOLD (Global Oriented Local Distribution Learning) is proposed. GOLD first finely explores the potential incomplete local cluster distributions of clients, then uploads the distribution summarization to the server for global fusion, and finally performs local cluster enhancement under the guidance of the global distribution. Extensive experiments, including significance tests, ablation studies, scalability evaluations, qualitative results, etc., have been conducted to show the superiority of GOLD.

Figures

Figures reproduced from arXiv: 2601.17512 by Haijun Zhang, Sen Feng, Shenghong Cai, Yiqun Zhang, Yuzhu Ji, Zihua Yang.

Figure 1
Figure 1. Figure 1: Non-ICD phenomenon and its impact on Federated Clustering (FC) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of GOLD. Each client first employs FCPL to extract compact local subclusters from Non-ICD data, then transmits only the corresponding [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Federated clustering performance comparison under different Non [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance comparison under different numbers of clients. [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Single-granular representation ablation of GOLD, evaluated by 4 [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Federated clustering process of GOLD in a real-world scenario. (a) visualizes the US dataset. (b) - (d) illustrate the process of cluster exploration on [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

56 extracted references · 3 linked inside Pith

  1. [1]

    Federated machine learning: Concept and applications,

    Q. Yang, Y . Liu, T. Chenet al., “Federated machine learning: Concept and applications,”ACM Transactions on Intelligent Systems and Tech- nology, vol. 10, pp. 1–19, 2019

    2019

  2. [2]

    FedDual: Pair-wise gossip helps federated learning in large decentralized networks,

    Q. Chen, Z. Wang, H. Wanget al., “FedDual: Pair-wise gossip helps federated learning in large decentralized networks,”IEEE Transactions on Information Forensics and Security, vol. 18, pp. 335–350, 2022

    2022

  3. [3]

    FedForgery: Generalized face forgery detection with residual federated learning,

    D. Liu, Z. Dang, C. Penget al., “FedForgery: Generalized face forgery detection with residual federated learning,”IEEE Transactions on Infor- mation Forensics and Security, vol. 18, pp. 1–15, 2023

    2023

  4. [4]

    Federated continual learning via knowledge fusion: A survey,

    X. Yang, H. Yu, X. Gaoet al., “Federated continual learning via knowledge fusion: A survey,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, pp. 1–18, 2024

    2024

  5. [5]

    PFLF: Privacy-preserving federated learning framework for edge computing,

    H. Zhou, G. Yang, H. Daiet al., “PFLF: Privacy-preserving federated learning framework for edge computing,”IEEE Transactions on Infor- mation Forensics and Security, vol. 17, pp. 1905–1918, 2022

    1905

  6. [6]

    Prototype similarity distillation for communication-efficient federated unsupervised representation learn- ing,

    C. Zhang, Y . Xie, T. Chenet al., “Prototype similarity distillation for communication-efficient federated unsupervised representation learn- ing,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, pp. 1–15, 2024

    2024

  7. [7]

    Privacy-preserving federated learning with malicious clients and honest-but-curious servers,

    J. Le, D. Zhang, X. Leiet al., “Privacy-preserving federated learning with malicious clients and honest-but-curious servers,”IEEE Transac- tions on Information Forensics and Security, vol. 18, pp. 1–15, 2023

    2023

  8. [8]

    Federated learning with hierarchical clustering of local updates to improve training on non-IID data,

    C. Briggs, Z. Fan, and P. Andras, “Federated learning with hierarchical clustering of local updates to improve training on non-IID data,” in Proceedings of the 20th International Joint Conference on Neural Networks, 2020, pp. 1–9

    2020

  9. [9]

    Adaptive clustering-based personalized federated learning framework for next POI recommendation with loca- tion noise,

    Z. Ye, X. Zhang, X. Chenet al., “Adaptive clustering-based personalized federated learning framework for next POI recommendation with loca- tion noise,”IEEE Transactions on Knowledge and Data Engineering, vol. 35, pp. 1–15, 2023

    2023

  10. [10]

    FedAPEN: Personalized cross- silo federated learning with adaptability to statistical heterogeneity,

    Z. Qin, S. Deng, M. Zhaoet al., “FedAPEN: Personalized cross- silo federated learning with adaptability to statistical heterogeneity,” in Proceedings of the 29th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2023, pp. 1954–1964

    2023

  11. [11]

    Lightweight privacy-preserving cross- cluster federated learning with heterogeneous data,

    Z. Chen, S. Yu, F. Chenet al., “Lightweight privacy-preserving cross- cluster federated learning with heterogeneous data,”IEEE Transactions on Information Forensics and Security, vol. 19, pp. 1–15, 2024

    2024

  12. [12]

    A survey on federated learning systems: Vision, hype and reality for data privacy and protection,

    Q. Li, Z. Wen, Z. Wuet al., “A survey on federated learning systems: Vision, hype and reality for data privacy and protection,”IEEE Trans- actions on Knowledge and Data Engineering, vol. 35, pp. 3347–3366, 2021

    2021

  13. [13]

    Federated learning on non-IID data silos: An experimental study,

    Q. Li, Y . Diao, Q. Chenet al., “Federated learning on non-IID data silos: An experimental study,” inProceedings of the 38th IEEE International Conference on Data Engineering, 2022, pp. 965–978

    2022

  14. [14]

    Federated multi-view spectral cluster- ing,

    H. Wang, A. Li, B. Shenet al., “Federated multi-view spectral cluster- ing,”IEEE Access, vol. 8, pp. 202 249–202 259, 2020

    2020

  15. [15]

    Federated learning approach for spectral clustering,

    E. Hern ´andez-Pereira, O. Fontenla-Romero, B. Guijarro-Berdin˜naset al., “Federated learning approach for spectral clustering,” inProceedings of the 29th European Symposium on Artificial Neural Networks, Compu- tational Intelligence and Machine Learning, 2021

    2021

  16. [16]

    Federated spectral clustering via secure similarity reconstruction,

    D. Qiao, C. Ding, and J. Fan, “Federated spectral clustering via secure similarity reconstruction,” inProceedings of the 38th International Con- ference on Neural Information Processing Systems, 2024, pp. 58 520– 58 555

    2024

  17. [17]

    Towards federated clustering: A federated fuzzyc-means algorithm (FFCM),

    M. Stallmann and A. Wilbik, “Towards federated clustering: A federated fuzzyc-means algorithm (FFCM),”arXiv:2201.07316, 2022

    Pith/arXiv arXiv 2022

  18. [18]

    Federated FCM: Clustering under privacy requirements,

    W. Pedrycz, “Federated FCM: Clustering under privacy requirements,” IEEE Transactions on Fuzzy Systems, vol. 30, pp. 3384–3388, 2021

    2021

  19. [19]

    K-means clustering with local d-privacy for privacy-preserving data analysis,

    M. Yang, I. Tjuawinata, and K.-Y . Lam, “K-means clustering with local d-privacy for privacy-preserving data analysis,”IEEE Transactions on Information Forensics and Security, vol. 17, pp. 2524–2537, 2022

    2022

  20. [20]

    Heterogeneity for the win: One- shot federated clustering,

    D. K. Dennis, T. Li, and V . Smith, “Heterogeneity for the win: One- shot federated clustering,” inProceedings of the 38th International Conference on Machine Learning, 2021, pp. 2611–2620

    2021

  21. [21]

    Fed-SC: One-shot federated subspace clustering over high-dimensional data,

    S. Xie, Y . Wu, K. Liaoet al., “Fed-SC: One-shot federated subspace clustering over high-dimensional data,” inProceedings of the 39th IEEE International Conference on Data Engineering, 2023, pp. 2905–2918

    2023

  22. [22]

    One-shot secure federated k- means clustering based on density cores,

    Y . Wang, W. Pang, D. Wanget al., “One-shot secure federated k- means clustering based on density cores,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, pp. 14 131–14 143, 2025

    2025

  23. [23]

    Machine unlearning of federated clus- ters,

    C. Pan, J. Sima, S. Prakashet al., “Machine unlearning of federated clus- ters,” inProceedings of the 11th International Conference on Learning Representations, 2023

    2023

  24. [24]

    Federated clustering via matrix factorization models: From model averaging to gradient sharing,

    S. Wang and T.-H. Chang, “Federated clustering via matrix factorization models: From model averaging to gradient sharing,”arXiv:2002.04930, 2020

    Pith/arXiv arXiv 2002

  25. [25]

    Federated k-means clustering: A novel edge AI-based approach for privacy preservation,

    H. H. Kumar, V . R. Karthik, and M. K. Nair, “Federated k-means clustering: A novel edge AI-based approach for privacy preservation,” inProceedings of the 9th IEEE International Conference on Cloud Computing in Emerging Markets, 2020, pp. 52–56

    2020

  26. [26]

    Federated fuzzy clustering for decentralized incomplete longitudinal behavioral data,

    H. Ngo, H. Fang, J. Rumbutet al., “Federated fuzzy clustering for decentralized incomplete longitudinal behavioral data,”IEEE Internet of Things Journal, vol. 11, pp. 14 657–14 670, 2023

    2023

  27. [27]

    One-shot federated clustering based on stable distance relationships,

    Y . Wang, W. Pang, and W. Pedrycz, “One-shot federated clustering based on stable distance relationships,”IEEE Transactions on Industrial Informatics, vol. 20, pp. 1–12, 2024

    2024

  28. [28]

    Differentially private federated clustering over non-IID data,

    Y . Li, S. Wang, C.-Y . Chiet al., “Differentially private federated clustering over non-IID data,”IEEE Internet of Things Journal, vol. 11, pp. 6705–6721, 2023

    2023

  29. [29]

    A comprehensive survey of privacy- preserving federated learning: A taxonomy, review, and future direc- tions,

    X. Yin, Y . Zhu, and J. Hu, “A comprehensive survey of privacy- preserving federated learning: A taxonomy, review, and future direc- tions,”ACM Computing Surveys, vol. 54, pp. 1–36, 2021

    2021

  30. [30]

    An efficient framework for clustered federated learning,

    A. Ghosh, J. Chung, D. Yinet al., “An efficient framework for clustered federated learning,” inProceedings of the 24th International Conference on Neural Information Processing Systems, 2020, pp. 19 586–19 597

    2020

  31. [31]

    Federated unsupervised clus- tering with generative models,

    J. Chung, K. Lee, and K. Ramchandran, “Federated unsupervised clus- tering with generative models,” inProceedings of the AAAI Workshop on Trustable, Verifiable and Auditable Federated Learning, 2022

    2022

  32. [32]

    Federated learning with soft cluster- ing,

    C. Li, G. Li, and P. K. Varshney, “Federated learning with soft cluster- ing,”IEEE Internet of Things Journal, vol. 9, pp. 7773–7782, 2021

    2021

  33. [33]

    FedUC: A unified clustering approach for hierarchical federated learning,

    Q. Ma, Y . Xu, H. Xuet al., “FedUC: A unified clustering approach for hierarchical federated learning,”IEEE Transactions on Mobile Comput- ing, vol. 23, pp. 9737–9756, 2024

    2024

  34. [34]

    Adaptive micro-partition and hierarchical merging for accurate mixed data clustering,

    Y . Zhang, R. Zou, Y . Zhanget al., “Adaptive micro-partition and hierarchical merging for accurate mixed data clustering,”Complex and Intelligent Systems, vol. 11, pp. 1–14, 2025

    2025

  35. [35]

    Sdenk: Unbiased subspace density-k-clustering,

    R. Zou, Y . Zhang, M. Zhao, Z. Tan, Y . Zhang, and Y .-m. Cheung, “Sdenk: Unbiased subspace density-k-clustering,”Neurocomputing, p. 131225, 2025

    2025

  36. [36]

    Federated hierarchical clus- tering with automatic selection of optimal cluster numbers,

    Y . Zhang, C. Qiu, X. Liao, and Y . Zhang, “Federated hierarchical clus- tering with automatic selection of optimal cluster numbers,”Information Sciences, p. 122957, 2025

    2025

  37. [37]

    Significance-based decision tree for interpretable categorical data clustering,

    L. Hu, M. Jiang, X. Liuet al., “Significance-based decision tree for interpretable categorical data clustering,”Information Sciences, vol. 690, p. 121588, 2025

    2025

  38. [38]

    Asynchronous federated clustering with unknown number of clusters,

    Y . Zhang, Y . Zhang, Y . Luet al., “Asynchronous federated clustering with unknown number of clusters,” inProceedings of the 39th AAAI Conference on Artificial Intelligence, 2025, pp. 22 695–22 703

    2025

  39. [39]

    Learning self-growth maps for fast and accurate imbalanced streaming data clustering,

    Y . Zhang, S. Feng, P. Wang, Z. Tan, X. Luo, Y . Ji, R. Zou, and Y .-M. Cheung, “Learning self-growth maps for fast and accurate imbalanced streaming data clustering,”IEEE Transactions on Neural Networks and Learning Systems, vol. 36, no. 9, pp. 16 049–16 061, 2025

    2025

  40. [40]

    Robust categorical data clustering guided by multi-granular competitive learning,

    S. Cai, Y . Zhang, X. Luoet al., “Robust categorical data clustering guided by multi-granular competitive learning,” inProceedings of the 44th IEEE International Conference on Distributed Computing Systems, 2024, pp. 288–299

    2024

  41. [41]

    Multimodal generative learning utilizing jensen-shannon-divergence,

    T. Sutter, I. Daunhawer, and J. V ogt, “Multimodal generative learning utilizing jensen-shannon-divergence,” inProceedings of the 24th Inter- national Conference on Neural Information Processing Systems, 2020, pp. 6100–6110. IEEE INTERNET OF THINGS JOURNAL, JANUARY 2026 15

    2020

  42. [42]

    Fast & accurate Gaussian kernel density estimation,

    J. Heer, “Fast & accurate Gaussian kernel density estimation,” in Proceedings of the 32nd IEEE Visualization Conference, 2021, pp. 11– 15

    2021

  43. [43]

    Comparing kullback-leibler di- vergence and mean squared error loss in knowledge distillation,

    T. Kim, J. Oh, N. Kimet al., “Comparing kullback-leibler di- vergence and mean squared error loss in knowledge distillation,” arXiv:2105.08919, 2021

    Pith/arXiv arXiv 2021

  44. [44]

    Maximum weighted likelihood via rival penalized EM for density mixture clustering with automatic model selection,

    Y .-M. Cheung, “Maximum weighted likelihood via rival penalized EM for density mixture clustering with automatic model selection,”IEEE Transactions on Knowledge and Data Engineering, vol. 17, pp. 750– 761, 2005

    2005

  45. [45]

    Spatio-temporal diffusion point processes,

    Y . Yuan, J. Ding, C. Shaoet al., “Spatio-temporal diffusion point processes,” inProceedings of the 29th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2023, pp. 3173– 3184

    2023

  46. [46]

    Anomaly detection with score distribution discrimination,

    M. Jiang, S. Han, and H. Huang, “Anomaly detection with score distribution discrimination,” inProceedings of the 29th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2023, pp. 984–996

    2023

  47. [47]

    Inter-node hellinger distance based decision tree,

    P. S. Akash, M. E. Kadir, A. A. Aliet al., “Inter-node hellinger distance based decision tree,” inProceedings of the 28th International Joint Conference on Artificial Intelligence, 2019, pp. 1967–1973

    2019

  48. [48]

    Detecting V oIP floods using the hellinger distance,

    H. Sengar, H. Wang, D. Wijesekeraet al., “Detecting V oIP floods using the hellinger distance,”IEEE Transactions on Parallel and Distributed Systems, vol. 19, pp. 794–805, 2008

    2008

  49. [49]

    A survey on homomorphic encryption schemes: Theory and implementation,

    A. Acar, H. Aksu, A. S. Uluagacet al., “A survey on homomorphic encryption schemes: Theory and implementation,”ACM Computing Surveys, vol. 51, pp. 1–35, 2018

    2018

  50. [50]

    Federated learning with differential privacy: Algorithms and performance analysis,

    K. Wei, J. Li, M. Dinget al., “Federated learning with differential privacy: Algorithms and performance analysis,”IEEE Transactions on Information Forensics and Security, vol. 15, pp. 3454–3469, 2020

    2020

  51. [51]

    Automated variable weighting in k-means type clustering,

    J. Z. Huang, M. K. Ng, H. Rong, and Z. Li, “Automated variable weighting in k-means type clustering,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 27, pp. 657–668, 2005

    2005

  52. [52]

    Comparing partitions,

    L. Hubert and P. Arabie, “Comparing partitions,”Journal of Classifica- tion, vol. 2, pp. 193–218, 1985

    1985

  53. [53]

    Cluster ensembles: A knowledge reuse frame- work for combining multiple partitions,

    A. Strehl and J. Ghosh, “Cluster ensembles: A knowledge reuse frame- work for combining multiple partitions,”Journal of Machine Learning Research, vol. 3, pp. 583–617, 2002

    2002

  54. [54]

    Silhouettes: A graphical aid to the interpretation and validation of cluster analysis,

    P. J. Rousseeuw, “Silhouettes: A graphical aid to the interpretation and validation of cluster analysis,”Journal of Computational and Applied Mathematics, vol. 20, pp. 53–65, 1987

    1987

  55. [55]

    A dendrite method for cluster analysis,

    T. Cali ´nski and J. Harabasz, “A dendrite method for cluster analysis,” Communications in Statistics: Theory and Methods, vol. 3, pp. 1–27, 1974. Yiqun Zhang(Senior Member, IEEE) received the B.Eng. degree from South China University of Technology, Guangzhou, China, in 2013, and the M.S. and Ph.D. degrees from Hong Kong Baptist University, Hong Kong SAR,...

    1974

  56. [2023]

    AR(ST)” row summarizes the average rank of each method across all datasets, and the symbol “(+)

    He is currently pursuing a master’s de- gree with the School of Computer Science and Technology, Guangdong University of Technology, Guangzhou, China. He has published a series of research works in reputable journals and confer- ences, including TNNLS, ECAI, and ICASSP. His current research interests include unsupervised ma- chine learning, ensemble clust...

    2012