This is the repository for paper V3H: View Variation and View Heredity for Incomplete Multi-view Clustering published by IEEE Transactions on Artificial Intelligence (TAI) by Xiang Fang, Yuchong Hu, Pan Zhou, and Dapeng Oliver Wu. Both IEEE Version and arXiv Version are available.
Multi-view clustering has wide applications in many real-world applications. In these applications, original image data often contain missing instances. In this repository, we implement a novel approach View Variation and View Heredity (V3H) for incomplete multi-view clustering.
We conduct extensive experiments on fifteen real-world datasets, and experimental results demonstrate V3H's superior advantages over other state-of-the-art clustering algorithms. The codes of the compared methods can be found on the authors' claimed websites.
.
├── run_V3H.m # DEMO file of V3H
├── V3H.m # core function of V3H
├── YaleB.mat # data mat files
├── splitDigitData.m # construction of incomplete multi-view data
├── solveF.m # the initialization of F
├── NormalizeFea.m # regularization of data
├── ClusteringMeasure.m # clustering performance
└── constructW.m, EuDist2.m, L2_distance_1.m, and readsparse.m # intermediate functions
MATLAB R2020a, Windows 10, 3.30 GHz E3-1225 CPU, and 64 GB main memory.
-
Install the MATLAB. The scripts have been verified in Matlab 2020a.
-
Download this repository via git
git clone https://github.com/ZeusDavide/TAI_V3H.git
or download the zip file manually.
-
Get multi-view dataset: We provide the YaleB dataset "YaleB.mat" in this repository as an example (you can download it from Google Drive). For the other datasets in the experiments, please refer to the corresponding links or articles.
-
Add the root folder to the Matlab path before running the scripts.
To reproduce the experimental results in Section V-D of the paper, we need to run the scripts run_V3H.m
.
- For
$\eta$ , we set$\eta=10^{-3}$ (i.e., relatively small$\eta$ ) for $||\bm{M}||{\eta}$ and $\tau=10^{-2}$ (i.e., relatively small $\tau$) for $||\bm{E}^{(v)}||{\tau}$. - In general, increasing iteration number
iter
will promote the clustering performance and consume more time. We recommend its maximum value is 30.
If you use this code please cite:
@ARTICLE{fangv3h2020,
author={Fang, Xiang and Hu, Yuchong and Zhou, Pan and Wu, Dapeng Oliver},
journal={IEEE Transactions on Artificial Intelligence},
title={V$^3$H: View Variation and View Heredity for Incomplete Multiview Clustering},
year={2020},
volume={1},
number={3},
pages={233-247},
doi={10.1109/TAI.2021.3052425}}