Dual Contrast-Driven Deep Multi-View Clustering

This repo contains the code and data associated with our DCMVC accepted by IEEE Transactions on Image Processing 2024.

Framework

The overall framework of the proposed DCMVC within an Expectation-Maximization framework. The framework includes: (a) View-specific Autoencoders and Adaptive Feature Fusion Module, which extracts high-level features and fuses them into consensus representations; (b) Dynamic Cluster Diffusion Module, enhancing inter-cluster separation by maximizing the distance between clusters; (c) Reliable Neighbor-guided Positive Alignment Module, improving within-cluster compactness using a pseudo-label and nearest neighbor structure-driven contrastive learning; (d) Clustering-friendly Structure, ensuring well-separated and compact clusters.

Requirements

hdf5storage==0.1.19

matplotlib==3.5.3

numpy==1.20.1

scikit_learn==0.23.2

scipy==1.7.1

torch==1.8.1+cu111

Datasets & trained models

The Cora, ALOI-100, Hdigit, and Digit-Product datasets, along with the trained models for these datasets, can be downloaded from Google Drive or Baidu Cloud password: data.

Usage

Train a new model：

python train.py

Test the trained model:

python test.py

Acknowledgments

Work&Code takes inspiration from MFLVC, ProPos.

Citation

If you find our work beneficial to your research, please consider citing:

@ARTICLE{10648641,
  author={Cui, Jinrong and Li, Yuting and Huang, Han and Wen, Jie},
  journal={IEEE Transactions on Image Processing}, 
  title={Dual Contrast-Driven Deep Multi-View Clustering}, 
  year={2024},
  volume={33},
  number={},
  pages={4753-4764},
  keywords={Feature extraction;Contrastive learning;Reliability;Clustering methods;Task analysis;Data mining;Unsupervised learning;Multi-view clustering;deep clustering;representation learning;contrastive learning},
  doi={10.1109/TIP.2024.3444269}}