【作者】 张杰；

【导师】 费伦科；

【作者基本信息】 广东工业大学 ， 计算机技术（专业学位）， 2023， 硕士

【摘要】 随着人工智能、大数据、物联网等技术的迅速发展,数据的多样性、异构性、复杂性越来越突出,这给传统的数据分析和处理带来了很大的挑战。在这样的背景下,多视角聚类因可以同时利用多个视角的信息,从而更全面地描述数据成为了一个备受关注的研究领域。但是在真实的应用场景中往往会因为一些不可控的环境和设备因素,导致数据缺失,这会给数据分析和处理带来困难。特别是在多视角数据中,缺失任何一个视角的信息都可能对最终的聚类结果造成影响。因此,找到一个能降低缺失造成性能下降的有效手段变成亟待解决的问题。现有的不完备多视角聚类方法虽然取得了不错的性能,但还是存在着一些缺点:在使用图正则项保留相似关系的方法中,对于缺失的信息往往简单的用零或者平均值填充,所以导致图正则项无法发挥合理的作用;在基于图的方法的共识图融合过程中,分配权重的方式往往是整个视角共用相同的权重,这无疑会导致视角间一致信息的损失。针对以上发现的两个缺点,本文将分别从恢复完整的图正则项以及设计更加合理的权重分配方式两个角度设计高效的不完备缺失多视角聚类算法,主要工作包括如下内容:(1)在多视角分析中,现有的聚类方法通常需要数据集中的数据都具有完整的视角信息,然而在真实应用场景下,样本在某些视角下往往是缺失的,这就造成现有的方法无法适应现实所需,针对这种情况本文设计了一种能够处理缺失问题的不完备多视角聚类模型。首先,使用矩阵分解技术学习各视角共识的表示矩阵。同时,为了充分利用缺失数据对以及现有数据对的相似信息,恢复了各视图完整的相似度矩阵。最后,为了保留数据的局部几何特征,使用图正则化约束共识表示矩阵。实验结果证明,该模型在多个数据集上均达到较高的聚类性能。(2)现有的基于图的不完整的多视角聚类方法会将不同的权重分配给不同的视角,以学习多视角的共识图。这种权重分配方式虽然降低了低质量图中噪点信息带来的负面影响,但同时也导致该视图中的正确信息无法发挥应有的作用。与现有的视图级分配权重方式不同,本文提出了一种简单有效的实例级权重分配方法用于各视角相似度图的融合。具体来说,首先使用可用视角的相似度信息来恢复缺失视角的相似度图。然后,给相似度图中的实例自适应的分配给不同的权重,以减小噪声的负面影响。最后,结合图融合和秩约束,可以为不完整的多视角数据学习到新的共识表示。该方法在多个多视角数据集上进行实验,实验结果表明与视图级别权重分配的方法相比,实例级别权重分配具有更好的聚类性能。综上所述,本文将从恢复完整的图正则项以及设计更加合理的权重分配方式两个角度对不完备多视角聚类进行研究设计模型,并在常用的多视角数据集上进行了大量实验,实验结果表明本文提出的两个模型能有效降低缺失带来的性能损失,并在三个性能指标上有更好的表现。更多还原

【Abstract】 With the rapid development of technologies such as artificial intelligence,big data,and the Internet of Things,the diversity,heterogeneity,and complexity of data have become increasingly obvious,posing great challenges to traditional data analysis and processing.In this context,multi-view clustering has attracted considerable attention as it can simultaneously utilize information from multiple views to provide a more comprehensive description of data.However,in real-world scenarios,data missing is often caused by uncontrollable environmental and equipment factors,which can pose difficulties for data analysis and processing.Especially in multi-view data,the missing information in any view can affect the final clustering results.Therefore,finding effective approaches to reduce the impact of missing data on the clustering performance has become an urgent issue.Although existing incomplete multi-view clustering methods have achieved good results,they still have some limitations.For example,in methods that use graph regularization to preserve similarity relationships,missing information is often filled with zeros or average values,resulting in a failure of the graph regularization method.In the consensus graph fusion process based on graph-based methods,the way of assigning weights often involves sharing the same weight across all views,leading to a loss of consistent information among views.This thesis proposes efficient incomplete multi-view clustering algorithms that address two limitations found in existing methods:the incomplete graph regularization problem and the need for a more reasonable weight allocation strategy for graph-based consensus fusion.we will design an efficient incomplete missing multi-view clustering algorithm from the perspectives of recovering complete graph regularization and designing a more reasonable weight distribution method.The main work includes the following:(1)In multi-view analysis,existing clustering methods often require that all data in the dataset have complete view information.However,in real-world applications,samples are often missing in some views,which makes existing methods unable to meet practical needs.To address this issue,we propose an incomplete multi-view clustering model that can handle missing data.Firstly,we use matrix factorization to learn the consensus representation matrix for each view.Additionally,to fully utilize similarity information from missing and existing data,we recover the complete similarity matrix for each view.Lastly,to preserve the local geometric features of the data,we use graph regularization to constrain the consensus representation matrix.Experimental results demonstrate that the proposed model achieves high clustering performance on multiple datasets.(2)Existing incomplete graph-based multi-view clustering methods assigned different weights to different views to learn a multi-view consensus graph.Although this weight distribution method reduces the negative impact of noise information in low-quality graphs,it also prevents the correct information in this view from playing its due role.Different from the existing view-level weight assignment methods,we propose a simple and effective instance-level weight assignment method for the fusion of similarity graphs from each view.Specifically,we first use the similarity information of available views to recover the similarity graph of missing views.Then,we adaptively assign different weights to instances in the similarity graph to reduce the negative impact of noise.Finally,combined with graph fusion and rank constraints,we can learn new consensus representations for incomplete multi-view data.The method is experimented on multiple multi-view datasets,and the experimental results show that the instance-level weight assignment has better clustering performance than the view-level weight assignment method.To sum up,this thesis will research and design models for incomplete multi-view clustering from the perspectives of recovering complete graph regularition and designing a more reasonable weight distribution method and conduct a large number of experiments on commonly used multi-view datasets.The experimental results show that the two models proposed in this thesis can effectively reduce the performance loss caused by missing and have better performance on the three performance indicators.更多还原