【作者】 孙超；

【导师】 袁运浩； 丁寿安；

【作者基本信息】 扬州大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 聚类作为机器学习领域的主要研究内容,其目标是通过某种相似性度量将对象数据聚成不同的簇,实现“物以类聚”。聚类的研究对象多种多样,其中一类近年来被广泛研究,即多视图数据。一般来说,单视图数据专注于描述对象的某些特定属性,不具有全面性。与其相比,多视图数据可以从多角度描述观察对象,相互之间具有较好的补充性,更加有利于聚类学习。另一方面,多视图数据也带来了新的挑战,例如,数据的高维度影响了聚类算法的学习性能。因此,如何将原始高维数据有效地映射到低维子空间,以提高多视图聚类算法的性能,是一个非常值得研究的问题。本文主要研究基于多视图深度典型相关分析的聚类算法及其应用,其主要研究内容和工作如下:(1)提出一种新的多视图子空间聚类算法(MFDCCA-SC)。该算法利用分数阶思想重构视图数据矩阵,以此降低原始数据中噪声造成的负面影响,并通过多视图深度典型相关分析学习多视图数据的低维公共子空间,利用学到的子空间非线性表示对数据进行聚类,将其划分为不同的簇。在COIL-100、F-MNIST和MNIST数据集上的实验结果验证了 MFDCCA-SC算法的有效性。(2)提出一种有监督的多视图子空间聚类算法(LMFDCCA-SC)。该算法利用数据集的标签信息构建标签嵌入散布矩阵作为视图内协方差矩阵,降低同类样本之间的离散度,减少视图协方差矩阵与真实值的偏差。在F-MNIST和MNIST两个数据集上的实验结果表明,LMFDCCA-SC算法比传统方法更加有效。(3)设计并开发基于深度相关表示的图像聚类分析系统。系统将上述提出的算法应用于实际,其主要部分包括图像预处理、多视图构建、子空间特征表示、聚类分析可视化等模块。在真实图像集上的测试证明了系统的可行性和实用性。更多还原

【Abstract】 Clustering is the main research content in the field of machine learning.Its goal is to cluster different object data into different clusters by some measure of similarity,achieving"like attracts like".Clustering has a variety of research objects,one of which has been widely studied in recent years,namely multi-view data.Generally speaking,single-view data focus on describing some specific properties of objects and are not comprehensive.In contrast,multi-view data can describe the observed objects from multiple angles,which are complementary to each other and more conducive to clustering learning.On the other hand,multi-view data also confront new challenges,such as the high dimension of data that affects the learning performance of clustering algorithms.To improve the performance of multi-view clustering algorithms,how to map the original high-dimensional data to a low-dimensional subspace effectively is very worth studying.This paper mainly studies clustering algorithms and applications based on multi-view deep canonical correlations.The main research contents and work are as follows:(1)A new multi-view subspace clustering algorithm(MFDCCA-SC)is proposed.The algorithm reconstructs the view data matrix by using the fractional order idea to reduce the negative impact caused by noise in the original data,and learns the low-dimensional common subspace of the multi-view data through multi-view depth canonical correlation analysis,clusters the data using the learned non-linear representation,and divides it into different clusters.The effectiveness of the MFDCCA-SC algorithm is validated on COIL-100,FMNIST and MNIST datasets.(2)A supervised multi-view subspace clustering algorithm(LMFDCCA-SC)is proposed.This algorithm complements the problem that the MFDCCA-SC algorithm cannot learn label information.Using the label information of the dataset to construct the label embedding scatter matrix as the intra-view covariance matrix,to reduce the dispersion between similar samples,it can effectively reduce the deviation between the view covariance matrix and the true value.The experimental results on two datasets,F-MNIST and MNIST,show that the LMFDCCASC algorithm is more effective than traditional methods.(3)Design and develop a cluster analysis system based on depth correlation representation,mainly for cluster analysis of image information.The system uses several algorithms proposed in this paper for practical applications,including a preprocessing module,a multi-view building module,a subspace feature representation module,and a cluster analysis visualization module.Tests on real image sets demonstrate the feasibility and practicality of the system.更多还原