Deep Multi-View Subspace Clustering with Anchor Graph

News

• [10.10] Our code is organized and released!
• [5.11] Our paper is online now: https://arxiv.org/pdf/2305.06939.pdf.
• Deep Multi-View Subspace Clustering with Anchor Graph (DMCAG) was accepted by IJCAI2023.

Table of Contents

• Project Description
• Installation
• Usage
• Citation

Project Description

We propose a novel deep self-supervised model for MVSC. A unified target distribution is generated via spectral clustering which is more robust and can accurately guide the feature learning process. The target distribution and learned features are updated iteratively. To boost the model efficiency, we use anchor graph to construct the graph matrix. This strategy can significantly reduce time complexity by sampling anchor points. Besides, we utilize contrastive learning on pseudo-labels to alleviate the conflict between the consistency objective and the reconstruction objective, thus consistent soft cluster assignments can be obtained among multiple views.

Installation

pip install -r requirements.txt

Usage

Prepare your dataset

The dataset would be a single .mat format file in a structure as the following:

dataset.mat
├── Y    1 * n
├── X1   n * d_1
├── X2   n * d_2
...   

Then place the file inside a folder named "data" and set up your dataset as follows in training file：

    if args.dataset ==  dataset_name:
        args.n_input = [d_1, ...]
        args.viewNumber =  v
        args.instanceNumber = n
        args.batch_size =  x
        args.n_clusters =  k
        args.arch= num of anchors
        args.gamma= q
        args.save_path = path/to/save/parameters

Train the model

To train the fully connected layer model using the specified arguments, you can run the following command:

python train_fc.py --lr 0.001 --n_z 10  --arch 50 --gamma 5  --dataset  dataset_name

You will get the clustering results as follows:

 Acc 0.8564, nmi 0.7321, ari 0.6459

Training a model with convolutional layers is the same as training a model with fully connected layers.

Citation

If you're using DMCAG in your research or applications, please cite using this BibTeX:

@article{cui2023deep,
  title={Deep Multi-View Subspace Clustering with Anchor Graph},
  author={Cui, Chenhang and Ren, Yazhou and Pu, Jingyu and Pu, Xiaorong and He, Lifang},
  journal={arXiv preprint arXiv:2305.06939},
  year={2023}
}