Research Of Graph Learning Method In Large-Scale Spectral Clustering Algorithm

With the development of information technology,human society generates massive amounts of data every day.How to mine effective information from these data as a decision-making support for various human social activities has become a key research direction in the information age.As an unsupervised learning technique,cluster analysis plays an important role in data mining and is often used to explore potential connections between unlabeled data samples.Spectral clustering algorithm has received extensive attention due to its good performance on real-world data sets,and the combination of bipartite graph ideas enables it to quickly process large-scale data sets.However,for the existing bipartite graph learning methods,some of them can only capture the local structure information in the data sample space,and some can only capture the global structure information in the data sample space.These two kinds of information are equally important,and the absence of either will lead to suboptimal clustering results.In order to solve the above problems,this thesis proposes two graph learning methods that can be applied to large-scale spectral clustering for single-view data and multi-view data respectively.The proposed method can simultaneously capture the local and global structural information in the data sample space.,thereby improving the clustering performance.The main work of this thesis is as follows:(1)For single-view data,this thesis proposes a structured comprehensive bipartite graph learning method.In this method,when learning the similarity between data samples,the adaptive neighborhood method that can only capture local structure information and the minimum self-representation error method that can only capture global structure information are combined into a joint optimization problem.The combination of these two methods will promote the comprehensive capture of local and global structural information in the data sample space during the entire learning process.Therefore,the bipartite graph learned based on this method will contain complete structural information in the data sample space.Thus enhancing the final clustering effect.At the same time,a Laplacian rank constraint is added to the joint optimization problem to ensure that the final learned bipartite graph directly has the structure expected by the clustering objective.Ablation experiments and comparative experiments on real datasets prove that this method is superior to other similar methods.(2)For multi-view data,this thesis proposes a multi-view fusion bipartite graph learning method.In the proposed method,the idea of joint optimization in structured comprehensive bipartite graph learning is applied to the learning of single-view similarity matrices to ensure the comprehensive capture of local and global structural information in the data sample space.Considering that the special information contained in different perspectives is not equal in importance,this method adds weight parameters for each perspective in the fusion process of the multi-view similarity matrix,and gives a strategy for its adaptive learning.At the same time,Laplacian rank constraints are also added to the fusion process to ensure that the fused bipartite graph has an ideal clustering structure.In order to promote the full use of information from each perspective,the single-view similarity matrix learning and multi-view similarity matrix fusion in this method are carried out jointly.A large number of experiments on real multi-view datasets prove that this method can achieve better clustering effect than other similar multi-view methods.