摘要
为了解决当前多视角图聚类方法依赖于谱分解而导致计算量高且超参数调优过程繁琐的问题,基于离散指示矩阵优化的方式,提出了一种无超参数的多视角图聚类方法。为了避免谱分解的连续松弛,采用一种超集群策略来高效区分数据的离散集群关系。在优化层面,改进了传统的坐标下降方法,以降低计算复杂度以及实现离散指示矩阵的快速优化。在物体图像、人脸图像、文本数据、手写字体图像中进行了算法性能验证。结果表明:相比于最近常用的多视角图聚类方法,所提方法在聚类精度和运行效率方面具有明显优势。
To solve problems of current graph-based multi-view clustering methods,such as big computational load and complex hyperparameter tuning process caused by overdependence on spectral decomposition,a high efficient graph-based multi-view clustering method with no hyperparameter was proposed by introducing discrete indicator matrix optimization.To avoid the continuous relaxation of spectral decomposition,a supercluster strategy was adopted to efficiently distinguish the discrete cluster relationships of data.For optimization,the traditional coordinate de-scent method was improved,which reduced the computational complexity,to realize the fast optimization of discrete indicator matrix.Additionally,the proposed method was verified in object images,face images,gene images and handwriting images.Results showed that the proposed method has obvious advantages in clustering accuracy and running efficiency compared with other graph-based multi-view clustering methods.
作者
王念
张炜
崔智高
苏延召
姜柯
李爱华
WANG Nian;ZHANG Wei;CUI Zhigao;SU Yanzhao;JIANG Ke;LI Aihua(Rocket Force University of Engineering,Xi’an 710025,Shaanxi)
出处
《火箭军工程大学学报》
2024年第3期51-59,共9页
Journal of Rocket Force University of Engineering
基金
陕西省自然科学基础研究计划(2023-JC-YB-501)。
关键词
多视角图聚类
离散指示矩阵
超集群策略
坐标下降
谱分解
graph-based multi-view clustering
discrete indicator matrix
supercluster strategy
coordinate descent
spectral decomposition
