Research On Multi-view Clustering Method Based On Dynamic Neighbor Learning

Clustering is a commonly used data analysis method in the field of machine learning,which has been widely used in practical applications such as information retrieval,medical diagnosis and anomaly detection.In the era of big data,data is generated from different sources or observed from different views,these data are called multi-view data.Compared with traditional data that describes objects from a single view,multi-view data is semantically richer and more useful,but more complex.In recent years,multiview clustering has become a research hotspot due to the traditional clustering algorithms can not deal with this kind of data.Multi-view clustering aims to use the consistency and complementarity information of multi-view data to divide data samples into several clusters.In order to make full use of the consistency and complementarity information between different views,seek the common representation of multi-view data,accurately discover the internal pattern of data,and achieve the purpose of improving the clustering accuracy of multi-view data,this paper conducts the research based on the dynamic neighbor learning method,the main research contents and innovations are as follows:1.In view of the problems that some existing multi-view clustering algorithms treat each view equally,and for the graph constructed for each view,noise and outliers are difficult to effectively filter out,a multi-view spectral clustering algorithm based on dynamic neighbor sparse graph learning(ASGL)is proposed.In the ASGL model,firstly,the dynamic neighbor graph learning method is used to construct the similarity matrix of all views,which improves the robustness to noise and outliers.Secondly,by dynamically assigning the weight of each view and combined with the complementary information between the views,the basic category attributes among sample data are described more accurately.Finally,extensive experiments on seven standard datasets demonstrate the superiority of the algorithm.2.Aiming at the problems that traditional multi-view clustering algorithms only focus on shared information in multiple views while ignoring the unique information and highorder correlation of each view,we propose a multi-view clustering algorithm(ANLTSC)based on dynamic neighbor learning and low-rank tensor decomposition.Specifically,the ANLTSC algorithm first learns the similarity graph of a view-specific by dynamically assigning the best neighbors to each data point based on the local distance,which effectively captures the local structural information between sample data.Next,the transition probability matrix corresponding to each view is calculated.Then,a tensor containing the transition probability matrix of each view is constructed.A tensor nuclear norm based on tensor-singular value decomposition(t-SVD)is used to constrain the rank of the target tensor.By minimizing the nuclear norm of the tensor,we learn a tensor that contains the shared information of each view and has a high-order correlation.Finally,extensive experiments are carried out on seven standard datasets,and the experimental results show that ANLTSC has a good clustering effect.3.Aiming at the problems that the multi-view clustering algorithms based on tensor nuclear norm ignore the prior information that each singular value has a significant difference,which leads to the obvious degradation of clustering performance for noisy data such as illumination changes,a hyper-Laplacian regularized multi-view clustering with a weighted tensor nuclear norm algorithm(WHLR-MSC)is proposed.Specifically,firstly,WHLR-MSC stacks the subspace representation matrices of different views into a tensor.Next,the tensor constructed by using the weighted tensor nuclear norm constraint can obtain the class discrimination information of the sample distribution more accurately.Then,hyper-Laplacian graph regularization is used to capture the local geometry structure of the data embedded in the high-dimensional ambient space.Finally,extensive experiments on five standard image datasets validate the effectiveness of the proposed WHLR-MSC method.