✨DCMVC: Dual Contrastive Multi-view Clustering

Pengyuan Li¹, Dongxia Chang¹, Zisen Kong¹, Yiming Wang², Yao Zhao¹

¹Institute of Information Science, Beijing Jiaotong University, Beijing, China
²School of Computer Science, Nanjing University of Posts and Telecommunications, Nanjing, China

🔥 Our work has been accepted by Neurocomputing 2025!

Overview🔍

Figure 1. The framework of the proposed DCMVC.

Abstract - Multi-view clustering, which aims to divide data into different categories that are unsupervised with respect to information from different views, plays an important role in the field of computer vision. Contrastive learning is widely used in deep multi-view clustering methods to learn more discriminative representations. However, most existing multi-view clustering methods based on contrastive learning use only a single positive sample and do not fully utilize the category information in the learning process. To address the above issues, we propose a novel dual contrastive multi-view clustering (DCMVC) method, which uses pseudo-labels to refine the embedded features to make them more suitable for clustering tasks. Specifically, an inter-view correlation contrastive module is designed to learn more compact clustering assignments through a shared clustering prediction layer. Then, on the basis of the clustering predictions, we propose an intra-view consistency contrastive module, which dynamically selects the samples with the same pseudo label as positive samples and sets the other samples as negative samples to construct contrastive learning. The proposed model can alleviate the constraints of a single positive sample on contrastive learning by fully considering the latent category information to regularize the representation structure. Extensive experiments conducted on nine real datasets demonstrate the superiority of the proposed method.

Datasets📚

To comprehensively evaluate the performance of the proposed DCMVC method, we conducted experiments using nine publicly available multi-view datasets, as shown below.

Dataset	Samples	Views	View Dimensions	Clusters
MSRC_v1	210	6	1302/48/512/100/256/210	7
Synthetic3d	600	3	3/3/3	3
Handwritten	2000	6	216/76/64/6/240/47	10
UCI-Digit	2000	3	216/76/64	10
NoisyMNIST	20000	2	784/784	10
Fashion	10000	3	784/784/784	10
BDGP	2500	2	1750/79	5
Mfeat	2000	6	216/76/64/6/240/47	10
NUSWIDEOBJ	30000	5	65/226/145/74/129	31

Experimental Results🏆

Table 1. Clustering average results and standard deviations for nine multi-view datasets. The optimal and suboptimal results are highlighted in red and blue. The notation O/M signifies an out-of-memory error encountered during the training process.

Table 2. Ablation study on MSRC_v1, UCI-Digit, and NoisyMNIST, respectively.

Getting Started🚀

Data Preparation

The dataset should be organised as follows, taking MSRC_v1 as an example:

MSRC_v1
├── X
│   ├── X1
│   ├── X2
│   ├── X3
│   ├── ...
├── Y

Training and Evaluation

• To train the DCMVC, run: main.py. The prediction results obtained using the K-Means algorithm.

Cite our work📝

@article{li2025dcmvc,
  title={DCMVC: Dual contrastive multi-view clustering},
  author={Li, Pengyuan and Chang, Dongxia and Kong, Zisen and Wang, Yiming and Zhao, Yao},
  journal={Neurocomputing},
  volume={635},
  pages={129889},
  year={2025},
  publisher={Elsevier}
}

License📜

The source code is free for research and educational use only. Any commercial use should get formal permission first.