Communication-Efficient Federated Multi-View Clustering via View-Aware Quantization

Khadija Ardif, Meryeme Ayache, Nidal Nasser, Ouiame MarnissiarXiv 2026 · arXiv

TL;DR

Federated Multi-View Clustering (FedMVC) allows multiple parties to collaboratively cluster data without sharing raw information, but it is notoriously bandwidth-heavy. This paper introduces View-Aware Quantization (VAQ), a dynamic strategy that prioritizes communication resources for the most informative data views, slashing communication costs by up to 6.59x while often improving clustering accuracy.

What Problem Does This Paper Address?

In the world of Federated Learning (FL), "multi-view" data refers to situations where different clients hold different types of features (views) for the same set of individuals—think of one hospital holding genomic data while another holds medical imaging for the same patients. Federated Multi-View Clustering (FedMVC) aims to find a shared structure among these views without moving the data to a central server.

The bottleneck is communication overhead. FedMVC requires repeatedly exchanging high-dimensional latent representations and cluster centroids between clients and a central server. Standard compression techniques (like uniform 8-bit quantization) treat every view as equally important. However, in reality, some views are noisy or redundant, while others are critical for defining the clusters. Wasting precious bandwidth on low-quality views slows down the system and can even degrade the final clustering results.

Methods

The authors propose a View-Aware Quantization (VAQ) algorithm that adaptively allocates a "bit budget" to each view based on its contribution to the global clustering task.

The process happens in two main phases:

  1. Initialization Phase: Before the main training starts, each client calculates an initial importance score based on its local data's clustering compactness (how well-defined the local clusters are) and representation variability (how much information the features carry).
  2. Dynamic Phase: During iterative training, the server updates these importance scores by looking at:
  3. Compactness Loss: A measure of how well the current global centroids fit the local data.
  4. Centroid Drift: How much the local centroids change between rounds. A large drift suggests the view is actively contributing new information to the global model.
Figure 2: Illustration of the proposed view-aware importance estimation and adaptive bit allocation strategy. Each client estimates the importance of its local view based on clustering statistics, and the resulting score is used to dynamically de- termine the quantization precision for communication-efficient transmission.

Based on these scores, the server solves a constrained optimization problem to decide exactly how many bits (from 1 to 16) each client should use to quantize its next upload. High-importance views get more bits (higher precision), while low-importance views are compressed more aggressively.

Datasets

The researchers evaluated VAQ across five diverse benchmark datasets commonly used in multi-view learning:

  • MNIST-USPS: Handwritten digits from two different sources.
  • Scene: Images categorized into 15 scene types.
  • BDGP: A biological dataset involving gene expression and images.
  • HW (Handwritten): Six different feature sets for handwritten numerals.
  • Caltech-5V: A subset of the Caltech-101 object recognition dataset.

Key Insights and Ideas

  • Quantization as Regularization: One of the most surprising findings was that VAQ-compressed models sometimes outperformed the "None" (full-precision) baseline. The authors suggest that adaptive quantization acts as a form of implicit regularization, filtering out noise from less informative views that might otherwise "confuse" the global clustering consensus.
  • Role Reversal: The importance of a view isn't static. The authors observed cases where a view was unimportant early on but became critical in later rounds as the global structure emerged. VAQ’s dynamic updates successfully tracked these shifts, redistributing bits in real-time.
  • Plug-and-Play Generality: VAQ isn't a standalone clustering method; it’s a communication module. The authors demonstrated its effectiveness by plugging it into four different state-of-the-art FedMVC frameworks (ESFMC, FCUIF, FMCSC, and HFMVC) without needing to change their internal architectures.

Key Results

The quantitative results show a clear win-win for communication and performance:

  • Massive Bandwidth Savings: VAQ achieved a maximum communication reduction ratio of 6.59x (on the Caltech-5V dataset using the HFMVC method) compared to the full-precision baseline.
  • Superior Accuracy: In many scenarios, VAQ significantly outperformed uniform 8-bit quantization. For the FCUIF method on the BDGP dataset, VAQ improved accuracy by 28.92% over uniform quantization while using a smaller total bit budget.
  • Efficiency: The computational cost of calculating importance scores is negligible ($O(nKr)$), matching the complexity of the underlying clustering steps.

Limitations and Future Work

While VAQ is highly effective, the authors note a few areas for improvement:

  • Manual Budgeting: The total global bit budget ($B$) currently needs to be set manually for each dataset/method combination. The authors plan to automate this selection in future work.
  • Discrete Data: The algorithm currently focuses on quantizing continuous latent representations. The discrete assignment matrices (which indicate which sample belongs to which cluster) remain uncompressed, representing a potential area for further optimization.
  • Hyperparameter Sensitivity: While a weighting factor of $\alpha = 0.5$ worked well across all tests, very specific edge cases might require more tuning to balance "static" view quality versus "dynamic" update contribution.

Summary updated

Abstract

Federated multi-view clustering (FedMVC) has recently attracted increasing attention for its ability to exploit complementary clustering structures across multiple clients while preserving data privacy. However, existing FedMVC methods often suffer from high communication overhead due to the repeated transmission of high-dimensional representations, which limits their scalability in practical federated systems. Moreover, different views may contribute unequally to the clustering performance, calling for adaptive communication strategies. In this paper, we propose a view-aware quantization algorithm for communication-efficient FedMVC. By dynamically assigning quantization bits based on iteratively updated view importance scores, the proposed algorithm effectively balances clustering performance and communication efficiency under a global bit budget constraint. To validate its generality, we integrate our approach into four state-of-the-art FedMVC methods. Extensive experiments on multiple benchmark datasets demonstrate that view-aware quantization achieves up to 6.59x communication reduction while maintaining and in some cases surpassing baseline clustering performance, confirming the effectiveness and scalability of the proposed approach.
BibTeX
@misc{ardif2026communication,
  author = {Khadija Ardif and Meryeme Ayache and Nidal Nasser and Ouiame Marnissi},
  title = {{Communication-Efficient Federated Multi-View Clustering via View-Aware Quantization}},
  year = {2026},
  eprint = {2611.15270},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15270},
}
Figure