Organizations: STRS Lab, National Institute of Posts and Telecommunications INPT, Rabat, Morocco · College of Engineering & Advanced Computing, Alfaisal University, Riyadh, Saudi Arabia · College of Computing Mohammed VI, Polytechnic University, Ben Guerir, Morocco
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.
Figures & tables
Notation
Definition
V
Number of federated clients (views)
v
Client / view index
n
Number of shared samples
T
Total number of federated communication rounds
K
number of clusters
t
communication round index
Table 1: Summary of Main Notations
Figure 1: Example of a FedMVC Architecture Illustrating the Role of VAQ. Each client i encodes its local view Xi into a latent representation Zi , locally clustered to yield soft assignments Yi and prototypes Ci . Prior to transmission, a local quantizer compresses Zi in a view-dependent manner. The server dequantizes the received uploads to reconstruct Z1′,…,Zv′ , aggregates them into a unified representation Z , and performs global clustering to produce {CGlobal,YGlobal} , which are broadcast back to refine each client’s quantizer. The process iterates until convergence. *Quantization is optional and method-dependent.
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 determine the quantization precision for communication-efficient transmission.
Dataset
Samples
Views
Clusters
MNIST-USPS
5000
2
10
Scene
4485
3
15
BDGP
2500
2
5
HW
2000
6
10
Caltech-5V
1400
5
7
Table 2: Statistics of the benchmark multi-view clustering datasets used in our experiments, including the number of samples, views, and clusters.
Method
Scenario
BDGP
HW
MNIST-USPS
Scene-15
Caltech-5V
ACC
NMI
ARI
ACC
NMI
ARI
ACC
NMI
ARI
ACC
NMI
ARI
ACC
NMI
ARI
FMCSC
None
92.88
80.88
83.42
93.60
86.80
86.35
94.06
88.34
87.53
39.49
37.44
21.61
82.57
70.33
67.76
Uniform-Q
94.24
83.51
86.38
92.45
85.53
84.12
92.88
87.51
85.50
39.40
37.62
21.52
82.00
70.73
66.93
View-Aware-Q †
94.60
84.58
87.06
94.30
88.17
87.76
94.44
88.82
88.26
40.65
37.24
21.74
84.43
72.39
69.88
FCUIF
None
66.80
38.51
35.72
63.35
58.97
47.77
50.66
47.27
34.53
30.58
30.94
17.62
46.43
40.22
30.01
Uniform-Q
52.32
30.49
25.56
62.85
64.66
53.42
51.30
49.05
35.76
32.34
31.03
18.71
54.50
42.04
29.35
Table 3: Clustering Performance Comparison Under Different Compression Scenarios Across Four FedMVC Methods and Five Benchmark Datasets. Bold results indicate that View-Aware-Q achieves equal or superior performance compared to the baseline, while underlined values denote a marginal performance reduction. † For View-Aware-Q, the total bit budget satisfies B≤V×8 bits.
Method
Scenario
BDGP
HW
MNIST-USPS
Scene-15
Caltech-5V
Round
Total
Red.
Round
Total
Red.
Round
Total
Red.
Round
Total
Red.
Round
Total
Red.
FMCSC
None
113.48
624.12
—
799.1003
4395.0518
—
109.49
602.20
—
195.13
1073.23
—
676.28
3719.57
—
Uniform-Q
28.37
156.03
4.00×
199.78
1098.76
4.00×
27.37
150.55
4.00×
48.78
268.31
4.00×
169.07
929.89
4.00×
View-Aware-Q †
27.03
149.35
4.18×
191.59
1057.82
4.15×
27.47
151.02
3.99×
47.71
262.95
4.08×
164.83
908.69
4.09×
FCUIF
None
0.1931
1.7405
—
0.4635
3.2501
—
0.3862
3.4809
—
0.5196
2.0847
—
0.2704
1.8959
—
Uniform-Q
0.0501
0.4530
3.84×
0.1202
0.8469
3.84×
0.1001
0.9060
3.84×
0.1347
0.5452
3.83×
0.0701
0.4940
3.84×
Table 4: Communication Overhead Comparison Under Different Compression Scenarios Across Four FedMVC Methods and Five Benchmark Datasets. Round: per-round communication cost (MB); Total: total communication cost (MB); Red.: reduction ratio vs. baseline. Bold values indicate that View-Aware-Q achieves a strictly higher reduction ratio than Uniform-Q, while underlined values denote a negligible difference. † The total bit budget for View-Aware-Q corresponds to the same configuration as in Table 3 .
Figure 3: Best clustering accuracy (ACC) as a function of the total bit budget for the proposed view-aware quantization, compared to the baseline (no quantization) and uniform 8-bit quantization, on the Caltech-5V dataset using ESFMC. View-Aware-Q with B=24 bits achieves the best ACC (91.14%), surpassing both the baseline (90.86%) and Uniform-Q (90.57%), while reducing the communication cost by 6.39 × compared to the baseline.
Figure 4: VAQ analysis in FMCSC on the HW dataset ( V=6 , Btotal≤V×8=48 bits). At each federated round, FMCSC estimates the importance of each client view and adaptively redistributes the bit budget accordingly, in contrast to the fixed uniform int8 allocation.
Variant
ACC (%)
NMI (%)
ARI (%)
Total Cost (MB)
Red.
VAQ-Uniform
92.45
85.53
84.12
1098.76
4.00×
VAQ-Static
93.05
86.21
85.27
1060.48
4.14×
VAQ-Full
94.30
88.17
87.76
1057.82
4.15×
VAQ-Uniform: uniform bit allocation ( bv=B/V ).
VAQ-Static: view-aware allocation with fixed initial importance scores.
VAQ-Full: proposed algorithm with dynamically updated importance scores.
Table 5: Ablation Study on the Contribution of Each Component of the Proposed VAQ Algorithm (FMCSC, HW Dataset). Bold values indicate the best result.
α
ACC (%)
NMI (%)
ARI (%)
Total Cost (MB)
0.0
93.80
87.21
86.79
1063.71
0.25
92.05
84.57
83.40
1057.10
0.5
94.65
88.59
88.51
1058.15
0.75
92.10
85.55
83.55
1066.88
1.0
91.14
83.48
81.66
1071.59
Table 6: Sensitivity Analysis of the Weighting Factor α on FMCSC evaluated on the HW Dataset. Bold values indicate the best result.
Multi-view clustering groups data by combining complementary representations while accounting for their shared structure, differences, and missing or noisy observations.
School of Computing and Information Technology, University of Wollongong, NSW, 2500, Australia · School of Software Engineering, Huazhong University of Science and Technology, Wuhan 430074, China