BGDR & CGDR for Multi-View Clustering

This repository contains the MATLAB implementation of the accepted paper:

Zhenyu Ma, Shengzhao Guo, Jingyu Wang, Feiping Nie and Xuelong Li, Scalable Graph Discrete Reconstruction for Efficient Multi-View Clustering, IEEE Transactions on Knowledge and Data Engineering (TKDE).

The code implements two efficient multi-view clustering methods:

• BGDR: Bipartite Graph Discrete Reconstruction
• CGDR: Compact Graph Discrete Reconstruction

Both methods directly learn discrete cluster indicators from multi-view data and use anchor-based graph construction for scalable clustering.

Folder Structure

.
+-- demo.m                  # Demo script for reproducing the paper settings
+-- MvC_BGDR.m              # Main function of BGDR
+-- MvC_CGDR.m              # Main function of CGDR
+-- Data/                   # Multi-view data sets in .mat format
+-- subfunc_BGDR/           # Shared utilities and BGDR subfunctions
`-- subfunc_CGDR/           # CGDR subfunctions

Each data file is expected to contain:

• X: a 1 x V cell array, where X{v} is the feature matrix of the v-th view
• label: the ground-truth label vector, used only for evaluation

Quick Start

Open MATLAB, enter this folder, and run:

demo

The script runs BGDR and CGDR on the included data sets using the parameter settings reported in the paper, and prints clustering metrics in the command window:

Dataset      Method  ACC     NMI    Purity   Precision   Recall  F-score  ARI   Time(s)
-----------------------------------------------------
MSRC_v1      BGDR    ...
MSRC_v1      CGDR    ...

Usage

BGDR:

[result,F,G,alpha,t,Obj,converge] = MvC_BGDR(X,label,k,h,initLabel,isNormal,maxIter);

CGDR:

[result,F,G,alpha,t,Obj,converge] = MvC_CGDR(X,label,k,h,initLabel,isNormal,maxIter);

Arguments:

• X: multi-view data cell array
• label: ground-truth labels for evaluation
• k: number of bipartite graph neighbors, denoted as r in the paper
• h: hierarchy depth for anchor generation, with anchor number m = 2^h
• initLabel: initialization method, default is 'N2HI'
• isNormal: whether to perform row-wise normalization, default is 1
• maxIter: maximum number of iterations, default is 30

Returned result contains seven clustering metrics:

[ACC, NMI, Purity, Precision, Recall, F-score, ARI]

Citation

If you find this code useful, please cite:

@ARTICLE{Ma2026GDR,
  author={Ma, Zhenyu and Guo, Shengzhao and Wang, Jingyu and Nie, Feiping and Li, Xuelong},
  journal={IEEE Transactions on Knowledge and Data Engineering}, 
  title={Scalable Graph Discrete Reconstruction for Efficient Multi-View Clustering}, 
  year={2026},
  volume={38},
  number={7},
  pages={4641-4657},
  doi={10.1109/TKDE.2026.3682510}
  }

Contact

For questions about the code or experiments, please contact zhenyu.ma@mail.nwpu.edu.cn