作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2024, Vol. 50 ›› Issue (5): 51-61. doi: 10.19678/j.issn.1000-3428.0067660

• 人工智能与模式识别 • 上一篇    下一篇

基于多样性与一致性的单步多视图聚类

胡傲然, 陈晓红   

  1. 南京航空航天大学数学学院, 江苏 南京 210016
  • 收稿日期:2023-05-18 修回日期:2023-07-21 发布日期:2024-05-14
  • 通讯作者: 陈晓红,E-mail:lyandcxh@nuaa.edu.cn E-mail:lyandcxh@nuaa.edu.cn
  • 基金资助:
    国家自然科学基金(11971231,12111530001)。

One-step Multi-view Clustering Based on Diversity and Consistency

HU Aoran, CHEN Xiaohong   

  1. College of Mathematics, Nanjing University of Aeronautics and Astronautics, Nanjing 210016, Jiangsu, China
  • Received:2023-05-18 Revised:2023-07-21 Published:2024-05-14
  • Contact: 陈晓红,E-mail:lyandcxh@nuaa.edu.cn E-mail:lyandcxh@nuaa.edu.cn

摘要: 随着数据采集技术的发展,多视图数据变得越来越常见。与单视图数据相比,多视图数据包含更丰富的信息,通常用一致性与多样性来刻画。现有基于图的多视图聚类方法大多只关注视图间的一致性信息,忽视了视图间的多样性信息,并且图的构建与聚类过程分离,从而影响聚类算法的效果。提出基于多样性与一致性的单步多视图聚类算法(OMCDC)。基于"距离较近的数据点成为邻居的可能性较大"这一先验知识构建各个视图的相似性图。不同于以往算法直接融合相似性图获得公共图,OMCDC将每个视图的相似性图分解为一致性图和多样性图,通过融合一致性图获得更具一致性的公共图。在此基础上,引入谱旋转,联合优化低维谱嵌入和聚类概率矩阵,将图学习和聚类融为一体,直接获得聚类结果。OMCDC充分利用了多视图数据的一致性信息与多样性信息,结合谱旋转实现了单步多视图聚类。实验结果表明,该算法在100L和HW2数据集上的聚类准确率分别为94.62%和99.30%,相比MVGL、AWP、MCGC等方法具有较优的聚类性能。

关键词: 多视图学习, 多视图聚类, 谱聚类, 谱旋转, 一致性, 多样性

Abstract: With the development of data collection technology, multi-view data have become increasingly common. Compared to single-view data, multi-view data contain richer information, which is usually characterized by consistency and diversity information. Most multi-view clustering methods based on graphs focus only on consistency information, neglect diversity information, and separate the construction of graphs from the clustering process, which may affect the clustering algorithm performance. This study proposes a One-step Multi-view Clustering algorithm based on Diversity and Consistency(OMCDC). It first constructs similarity graphs for each view based on the prior knowledge of ″data points with smaller distances are more likely to become neighbors.″ Second, unlike previous algorithms that directly fuse similarity graphs to obtain a common graph, this study decomposes the similarity graphs of each view into consistency and diversity graphs, and thereafter obtains a more consistent common graph by fusing the consistency graphs. Furthermore, spectral rotation is introduced to jointly optimize the low-dimensional spectral embedding and clustering probability matrix, integrating graph learning and clustering to obtain the clustering results directly. OMCDC fully uses the consistency and diversity information of multi-view data and combines spectral rotation to achieve one-step multi-view clustering. The clustering accuracies of this method on the 100Leaves(100L) and HandWritten digits2(HW2) datasets are 94.62% and 99.30%, respectively. Compared with Graph Learning for Multi-View clustering(MVGL), multi-view clustering via Adaptively Weighted Procrustes(AWP), and Multi-view Consensus Graph Clustering(MCGC), OMCDC achieves better clustering performance.

Key words: multi-view learning, multi-view clustering, spectral clustering, spectral rotation, consistency, diversity

中图分类号: