.png){kind=icon}
返回
Spectral type subspace clustering methods: multi-perspective analysis
.png){kind=icon}
PRE
AI
DOI：10.1007/s11042-023-16846-0.png)
.png){kind=icon}
原文求助.png){kind=icon}
分享.png){kind=icon}
收藏
摘要
En 中文
Founded on the premise that high-dimensional data can be characterized as data drawn from a union of several low-dimensional subspaces, subspace clustering has become famous due to the limitations of traditional clustering techniques such as k-means. Among the subspace clustering methods, spectral-based techniques have become increasingly popular in the last decade due to their potential to handle corrupt samples through the self-expressiveness property of data. It is, therefore, crucial to compare the spectral-based methods with each other. However, previous studies often analyze them from a single perspective, which does not tell the whole story. This paper presents an analysis of existing spectral-based methods from two perspectives: single-view and multi-view. Firstly, a detailed fundamental of subspace clustering is presented. Afterward, an overview of available techniques is provided from the two perspectives. Furthermore, we evaluate the clustering performances of current techniques on four datasets: UCI-Digits, Yale, COIL20 and Caltech101-07. In each paradigm, we first compare some existing approaches against each other and then investigate the improvement of the multi-view approaches over the single-view methods. From the results of different experiments, it was evident that not all multi-view methods outperform their single-view counterparts consistently. Surprisingly, in certain datasets, some single-view methods even outperformed some multi-view methods. To strengthen this comparison, we compare the similarity matrices of the different techniques. Finally, we highlight the challenges and recommend future research.
Keyword:
Subspace clustering
Low-rank representation
Sparse subspace clustering
Multiview clustering
期刊
IF:
3
论文数:
2.0W
被引数:
3.2W
机构
