acceptodds
Under review as a conference paper at ICLR 2027

Incremental Vertical Federated Multi-View Clustering with Asynchronous Client Arrival

Abstract

Existing vertical federated multi-view clustering methods generally assume that all clients are simultaneously available and that the client set is fixed. In practice, however, each client holds one view of the same samples, while clients may not be simultaneously valid and may join asynchronously over time. This setting exposes two limitations of current approaches. First, they typically require previous clients to participate again, which is infeasible when old clients are no longer available. Even when they remain available, such repeated participation still incurs extra communication and privacy costs. Second, given only the historical global consensus and one newly uploaded label matrix, existing methods lack an effective fusion mechanism. To address these issues, we propose an asynchronous incremental vertical federated multi-view clustering framework. Each arriving client constructs an anchor graph from its own view and learns a local label matrix, which is uploaded to the server. The server then collaborates with the current client to alternately update the global consensus and the local label matrix via an augmented Lagrangian formulation until convergence. Only the current client is involved, and the server retains only the latest consensus. Experiments on multiple datasets confirm the applicability and effectiveness of the proposed framework in incremental scenarios.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-view clustering, Federated Multi-View Clustering, Incremental Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_QunbOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies vertical federated multi-view clustering in an asynchronous incremental setting, where clients arrive sequentially and earlier clients may no longer participate. Each client performs local clustering and uploads only a label matrix. The server interacts only with the current client, alternating between updating the global consensus and the newly received label matrix. The two matrices are stacked into a third-order tensor and regularized by a tensor Schatten p-norm with dynamic weights, solved via augmented Lagrangian and projected gradient ascent, with a closed-form tensor thresholding step. Experiments on six benchmark datasets compare with centralized and synchronous federated methods using ACC, NMI, and Purity, and also analyze anchor rate, parameter sensitivity, and convergence. The work provides a practical incremental framework that avoids re-engaging previous clients and keeps server storage independent of the number of historical clients.

Critical strengths

1.The problem setting is practically meaningful and novel. The paper targets a realistic vertical federated multi-view clustering scenario where clients join sequentially and old clients may be offline. Most existing methods assume all clients are simultaneously available and require repeated participation, which is often unrealistic. The proposed framework avoids re-engaging previous clients, keeps server-side storage independent of the number of historical clients, and reduces repeated communication and exposure. This is a valuable contribution to federated clustering.

2.The method is lightweight and flexible. Clients upload only a label matrix, not raw features, gradients, or model parameters. This reduces server-side intrusion into local models and avoids direct transmission of feature-level representations. The framework also imposes no restriction on the local clustering model, allowing heterogeneous client-side algorithms. Experiments on six datasets show results comparable to, and in some cases better than, centralized and synchronous federated methods, which supports the feasibility and applicability of the incremental fusion idea.

Critical weaknesses
  1. The current experiments provide useful evidence of final clustering performance, but the incremental process itself could be illustrated more explicitly. If feasible, it would further strengthen the paper to show how the global consensus evolves as clients arrive sequentially and how the incremental update compares with a simple fusion baseline.
Questions for authors

1.The fusion in Eqs. (2) and (3) appears to require row-wise correspondence between the historical consensus and the newly uploaded label matrix. Could the authors clarify whether perfect sample alignment is assumed, i.e., identical sample IDs in the same row order, and whether this assumption holds throughout the experiments? 2. For practical deployment, could the authors clarify the communication cost of one incremental update upon a new client arrival and, if feasible, compare it with the cost of retraining a standard synchronous federated clustering model over all clients? 3. The paper reports parameter sensitivity but does not explain how the hyperparameters were selected in the main experiments. Could the authors clarify the search ranges and selection criteria, and whether the same tuning protocol was followed across all datasets?

1 comment
AuthorsOpenReview ↗

Response to Reviewers

We sincerely thank you for the constructive feedback. Your suggestions regarding the incremental process, sample alignment, communication cost, and hyperparameter selection are highly valuable for improving the practicality and rigor of our paper. We respond as follows:

Response to Critical Weaknesses

On the evolution of global consensus (Per-arrival curves): We fully agree. In the revision, we will add per-arrival curves to illustrate how the global consensus evolves and converges smoothly as new clients arrive sequentially.

On the comparison with a simple fusion baseline: We accept this suggestion. We will introduce a naive Simple Linear Averaging Fusion baseline for comparison. We will demonstrate that, compared to simple matrix averaging, our tensor Schatten-p fusion mechanism more effectively extracts cross-view complementary information and resists feature drift during the incremental process.

Response to Questions for Authors

Q1: On row-wise correspondence and sample alignment assumption A: This is a critical point. Yes, under our Vertical Federated setting, we assume perfect sample alignment, meaning all clients have identical sample IDs and maintain the same row order when uploading label matrices. This is a standard premise in vertical federated learning. All benchmark datasets used in our experiments (e.g., Yale, 3-sources, Sonar) were strictly aligned by sample ID and order during preprocessing. To eliminate ambiguity, we will explicitly state this assumption in the Problem Formulation section of the revision.

Q2: On communication cost of incremental update A: We objectively compare the communication cost of our method with standard synchronous federated methods. Single incremental update cost: When a new client arrives, the server and this client perform alternating iterations until convergence. Let T be the number of iterations (experiments show T is typically a single digit). Each round transmits the n-by-k label matrix H_v and consensus H_s. The total communication cost per update is O(T * n * k). Comparison with synchronous FL: Standard synchronous federated clustering requires all M clients to stay online for T global iterations, with a total cost of O(T * M * n * k). Our method's cost does not grow with the number of historical clients M, and it completely eliminates the need to recall historical clients, offering significant communication and privacy advantages in incremental scenarios.

Q3: On hyperparameter search ranges and selection criteria A: For absolute fairness in benchmarking, we adopted a unified Grid Search Protocol across all datasets. Search ranges: Anchor rate in [0.001, 1.0], client regularization weight λ_v in [0.0001, 0.01] (empirically, optimal values consistently fall within [0.0003, 0.008], validating our theoretical analysis on minimal weights), server tensor weight λ in [1, 12], and parameters p and p_1 in [0.1, 1.0]. Selection criteria: To explore the upper-bound performance, we adopted the Oracle evaluation paradigm using ground-truth labels to select the combination that maximizes clustering metrics (ACC/NMI). We will explicitly add these details to the revision.

Thank you again for your valuable suggestions. We will fully implement these additions in the revised manuscript.

Reviewer_quQZOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies vertical federated multi-view clustering in which clients arrive one at a time and never return, and fuses each client's anchor-graph soft label matrix from Eq. 1 with a stored consensus using a tensor Schatten-pp norm, proximal terms and adaptive weights solved by ADMM. On six small benchmarks against three centralized and six federated baselines, it reports for example 0.842 ACC on Yale and 0.533 ACC on NUS-WIDE in Tables 1 and 2.

Critical strengths

Clients that join one at a time and never return are a realistic setting that federated multi-view clustering has largely overlooked, and the server cost here does not grow with the number of past clients.

Critical weaknesses
  1. The server problem has no fixed objective, because Eq. 2 penalizes only the distance of HstH_s^{t} and HvtH_v^{t} to their previous iterates, so the stored consensus Hs0H_s^{0} serves only as a starting point and nothing anchors the fused result to the views it summarizes. The Eq. 25 weights are ratios of iterate-difference norms that become 0/0 at convergence even though Eq. 3 uses them to form the final consensus, and no convergence guarantee is given for the alternation with the client update of Eq. 14.

  2. Clustering sequentially arriving views with only a stored consensus is the continual multi-view clustering setting of CMVC, FCMVC-IV, ACMVC and AdaptCMVC, none of which is cited, and the server fusion combines the tensor stacking of indicator matrices from TensorFMVC with the Schatten-pp regularization of FFCMSP and FIMVC-TLRC. The framing also overstates the setting, since the "asynchronous" arrival in Sec. 2 is purely sequential with no concurrent clients or stale updates, and Sec. 3.1 claims privacy without a threat model or attack analysis although every round uploads an n×kn\times k soft assignment revealing record-level cluster membership.

  3. The GMC, UDBGL, FastMICE and TensorFMVC entries on five datasets are identical to the tables of TensorFMVC by Feng et al. 2024, whereas the proposed method is a single run chosen from per-dataset sweeps over five hyperparameters in Figs. 1 and 2, with no standard deviations. The incremental claim itself is untested, because all competitors are synchronous and only final-stage results appear, with no per-arrival curve, arrival-order study, synchronous run of the same fusion, or ablation of the proximal terms, tensor norm and adaptive weights.

Questions for authors

1.Were the baseline numbers rerun under your max-normalization and client splits or taken from the TensorFMVC paper, and how were the anchor rate, p1p_1, λv\lambda_v, pp and λ\lambda selected without ground-truth labels? Please also give nn, kk, the number of clients and the arrival order for each dataset.

2.What fixed objective does the server iteration of Eq. 15 minimize, and how are the Eq. 25 weights used in Eq. 3 defined once both iterate differences vanish? A convergence argument, or a term tying HsH_s to the stored Hs0H_s^{0}, would show whether the historical consensus is actually preserved.

3.Because DvD_v, the product of SvS_v with its transpose, is entrywise nonnegative, the trace term in Eq. 1 never exceeds the sum of all entries of DvD_v for a nonnegative row-stochastic HvH_v and reaches that bound when all samples share one cluster, so maximizing it favors collapse. What prevents this collapse as λv\lambda_v grows, and why maximize this term instead of minimizing a Laplacian smoothness term, as the "smoothness" description in Sec. 3.1 suggests?

1 comment
AuthorsOpenReview ↗

Response to Reviewers

We sincerely thank you for the highly rigorous and insightful review.

  1. Regarding CW1 & Q2

On the "term tying H_s to H_s^0": Our framework is online incremental learning; Eq. 2 is a dynamic online surrogate objective. The first penalty term in Eq. 2 precisely serves as the "term tying H_s to the stored H_s^0" you suggested (i.e., a dynamic proximal term). By restricting adjacent iteration deviations, it ensures historical consensus is smoothly transferred, preventing catastrophic forgetting.

On the 0/0 Singularity: This singularity does not occur in actual numerical computation. For theoretical rigor, we will introduce a very small smoothing constant in the revision to close this loophole.

On Convergence: While a strict global proof is intractable due to the non-convexity of the Schatten-p norm, numerical stability has a solid analytical foundation. We have proven that the subproblems for updating H_s and H_v strictly decompose into closed-form solutions via probability simplex projection, and the tensor T update also has an exact closed-form solution based on FFT. This alternating optimization design fundamentally ensures numerical stability. Fig. 3 empirically shows rapid convergence within a few iterations.

  1. Regarding CW2

Literature & Setting: We accept the criticism regarding the omission of CMVC et al. and will fully supplement the Related Work. We clarify that we target Vertical Federated (feature-isolated, unrecallable historical clients), distinct from centralized continual clustering. Our "theoretical upper-bound comparison" (approaching/surpassing centralized algorithms like GMC, FastMICE, TensorFMVC) logically validates its effectiveness.

On "Asynchronous": To distinguish from strict global synchronization, our "asynchronous" means the server updates incrementally without recalling historical clients. We will explicitly define this in the Introduction.

On Component Stacking: While tensor stacking and Schatten-p regularization are general foundational tools, conventional tensor fusion relies on the simultaneous presence of all views. To break this limitation, our core contribution lies in fundamentally reformulating the tensor fusion mechanism. We design a novel server-current client alternating optimization architecture specifically tailored to the asynchronous incremental vertical federated scenario. This architecture enables tensor fusion to operate solely based on historical consensus and local incremental information, without requiring access to global data. This constitutes a system-level architectural innovation, rather than mere component stacking.

On Privacy: We agree that the n x k soft assignment matrix poses leakage risks. Our contribution is design-level information dimensionality reduction. We will introduce an honest-but-curious threat model and state our framework is seamlessly compatible with standard Local Differential Privacy (LDP).

  1. Regarding CW3 & Q1

Baselines: We admit directly citing baseline results from TensorFMVC. To ensure absolute fairness, we will re-run the relevant baselines with identical preprocessing to verify robustness.

Single-Run: To strictly control variables and ensure reproducibility, a uniform random seed is fixed for all compared algorithms that involve stochastic processes.

Parameters: Hyperparameters were chosen via Oracle evaluation (using labels), a standard benchmarking paradigm to explore theoretical upper bounds. In unlabeled scenarios, unsupervised metrics (e.g., silhouette) can guide selection.

Dataset Details: Arrival order is natural (View 1 to View M). Parameters (n, k, Clients): Yale (165, 15, 2); 3-sources (169, 6, 3); Sonar (208, 2, 3); RGB-D (1449, 13, 2); VS (1594, 2, 4); NUS (30000, 31, 5).

Supplement: To directly validate the incremental advantage, we will add per-arrival curves comparing our method against a naive Simple Averaging baseline.

  1. Regarding Q3

We are grateful for your sharp mathematical insight. Your derivation that maximizing the trace term as lambda_v approaches infinity leads to collapse is completely correct. We retain our design based on:

Engineering Control: Regularization focuses on fine-tuning, not dominating. We strictly limit lambda_v to a small optimal range [0.0003, 0.008], ensuring the Schatten-p term absolutely dominates global clustering balance, directly avoiding the collapse limit.

Gradient Dynamics: Since we use Projected Gradient Ascent, the Laplacian gradient includes a penalty from the degree matrix (D_deg). Under a minimal lambda_v, the core optimization dynamics of both forms are highly consistent. However, the Laplacian's D_deg penalty mathematically penalizes assignment confidence (pushing toward a uniform distribution). In the loosely coupled vertical FL setting, this inherently weakens the discriminative sharpness of cluster assignments. Our trace term achieves local smoothness without penalizing assignment confidence. Thus, we retain it.

Thank you for your guidance.

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

In vertical federated clustering, when existing clients depart and new ones join, the shared partition must be updated without re-accessing the data from the previous clients. To address this scenario, the authors propose an approach that maintains a consensus label matrix on the server side, which is then exchanged exclusively with newly joined clients. Clients learn soft assignments via anchor graphs, while the server integrates historical consensus with current assignment results, employing tensor Schatten-p regularization and adaptive weighting, in an iterative process alternating with local client-side optimization until a stopping criterion is met. Experimental results across six datasets demonstrate that this method achieves competitive final clustering scores.

Critical strengths
  1. The authors address a useful operational constraint by updating a shared clustering result without contacting earlier clients and by retaining only the latest consensus, keeping persistent server storage independent of the number of previous clients for fixed sample and cluster counts.
  2. The reported final scores are promising on several datasets, including ACC improvements, which merit further investigation.
Critical weaknesses
  1. Equations (2) and (3) combine label matrices column by column without specifying cluster-label alignment or establishing invariance to independent column permutations, so even clients describing the same partition with different cluster indices may be fused inconsistently.

  2. Equation (2) penalizes changes from the previous iterate rather than a fixed historical reference, while the client separately optimizes Equation (1), leaving the joint objective unclear and the claim of a globally optimal consensus unsupported by the short convergence plots, with Equation (25) also undefined when both update norms vanish.

Questions for authors

How are cluster columns aligned before tensor fusion, and do the results remain unchanged when each client's label matrix is independently column-permuted without changing its partition?

2 comments
AuthorsOpenReview ↗

Response to Reviewers

We sincerely thank you for the highly rigorous and insightful review. Your concerns regarding column alignment, joint objective consistency, and formula definitions address core challenges in federated multi-view clustering. We fully agree and respond as follows:

  1. Regarding Column Alignment and Permutation Invariance (Eq. 2 & 3)

In our incremental setting, the server fuses exactly two matrices at a time: historical consensus H_s and new client H_v. Theoretically, independent matrices may have mismatched column indices. Our framework ensures stability via internal mechanisms without complex post-processing: (1) Anchor Consistency: As stated in Sec 3.1, all clients use identical anchor sampling and graph construction. Shared underlying semantics give H_s and H_v a strong implicit alignment tendency. (2) Tensor Schatten-p Penalty: Misaligned columns destroy the tensor's low-rank property. Minimizing the tensor Schatten-p norm in Eq. 2 imposes a massive penalty on misalignment. In the gradient dynamics, this strongly suppresses misaligned states, forcing structural consistency in the tensor space. We will add a brief theoretical explanation of this alignment penalty in Sec 3.2 of the revision.

  1. Regarding Eq. 2 Fixed Reference and Joint Objective

Eq. 2 is a dynamic online surrogate objective. Its first penalty term is the dynamic proximal term tying H_s to H_s^0, preventing catastrophic forgetting. Regarding the "unclear joint objective," we clarify: Eq. 1 and Eq. 2 are solved via alternating optimization. This follows the standard alternating optimization paradigm, utilizing block coordinate updates to seek a stable joint stationary point. On the "globally optimal consensus," we agree that a strict global proof is intractable in non-convex optimization. We clarify: the "consensus" refers to the stable fused solution upon convergence. We will revise the wording to "converged federated clustering consensus." Numerical stability is analytically grounded by exact closed-form solutions (simplex projection and FFT), with Fig. 3 empirically demonstrating rapid convergence.

  1. Regarding Eq. 25 Singularity

In practice, we avoid the 0/0 singularity via early stopping. For theoretical rigor, we will introduce a small smoothing constant epsilon to the denominator of Eq. 25 in the revision.

  1. Answer to the Question

Q: How are columns aligned before fusion? Do results change if permuted? A: The server processes only two matrices per step. Driven by anchor consistency and the tensor low-rank penalty, they exhibit a strong alignment tendency. Misalignment sharply increases the regularizer in Eq. 2, which gradient dynamics strongly suppress. Thus, even with independent column permutations of the new client's matrix, the model exhibits significant robustness and a strong convergence tendency towards alignment.

Thank you for your detailed response and for the time and effort devoted to considering my comments. I appreciate your careful engagement with the points raised and the clarifications you have provided. I also appreciate your willingness to revise the manuscript accordingly.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.