Research On Multi-view Clustering Algorithms Based On Adaptive Graph Construction

Clustering algorithm has always been a hot and important research field in machine learning,and has been widely applied in the field of image recognition,image processing,and message analysis.Nowadays,with the explosive growth of information,different descriptions of the same data may be widely divergent.Therefore,the single processing of data may not be the best method,and the errors caused will also make the clustering description of data inaccurate.In recent years,multi-view clustering method has become a hot research direction.Different from the previous single view clustering,clustering is the main means of the multiple views through the different views of the data structure of multiple similarity matrix,or by the express way to construct multiple views Shared subspace or independent subspace to fusion,etc.,finally structure similarity matrix or subspace for fusion clustering clustering results are obtained.However,most of the current multi-view learning clustering algorithms directly fuse the target matrix,and the effect in some practical experiments is not satisfactory,because the direct fusion process does not fully consider the difference between the similarity matrix of each view.Secondly,in describing the similarity between the data,the existing Euclidean distance or Gaussian distance is not necessarily the optimal solution.Thirdly,in terms of weight measurement of multi-view clustering algorithm,measuring only through similarity or measuring different views without considering weight will affect the final result of clustering.In the multi-view subspace clustering algorithm,although the subspace effectively solves the dimension disaster and ensures the data structure,the lack of in-depth exploration of the relationship between views,and the potential clustering number cannot be expressed only through the overall data structure,which leads to the deviation of clustering results for the multi-view clustering algorithm.In order to solve the above problems,this paper proposes three kinds of multi-view clustering algorithms based on adaptive composition from the perspectives of manifold adaptive composition,low-rank matrix decomposition and adaptive hypergraph construction,and improves the performance of multi-view clustering algorithm.The main research work and achievements are as follows:(1)A multi-view clustering algorithm based on manifold adaptive composition is proposed.Manifold distance is used to describe the manifold relationships between data elements,at the same time,by sharing cluster indicating the structure of the matrix to oversee the manifold similarity matrix,and then adaptively assigning different weights to multiple view similarity matrix of data fusion is a center manifold similarity matrix,the clustering result is acquired by spectral clustering directly.The algorithm uses the adaptive composition algorithm to construct the similarity matrix of multiple views to express the structural relations of views,and guarantees the structural characteristics of multiple similar matrices,thus effectively improving the performance of multi-view clustering algorithm.Experimental results on a large number of real data sets show that our proposed method has better advantages than other multi-view clustering algorithms.(2)A multi-view subspace clustering algorithm based on low rank sparse adaptive composition is proposed.In this algorithm,the subspace representation of multiple views is decomposed into potential clustering orthogonal projections through low-rank matrix decomposition,and at the same time,sparseness and low-rank properties are applied to ensure the local structure and global structure of data,and then the learned clustering orthogonal projections are fused into the central fusion matrix.Then the similarity matrix is constructed according to the center fusion matrix adaptively for clustering.Experimental results on a large number of real data sets demonstrate the effectiveness of our algorithm.(3)A multi-view clustering algorithm based on adaptive hypergraph construction is proposed.Based on the adaptive composition clustering algorithm,this algorithm uses the similarity relationship between views,constructs the similarity matrix of multiple views through the adaptive composition algorithm,and then according to the elements of multiple similarity matrix.Structure similarity matrix of similarity matrix,we call it a hypergraph,finally give the similarity matrix elements give different adaptive weight,this can be achieved in adaptive clustering effect of multiple views at the same time makes the similarity matrix is also considering the weights of relationship between clustering effect,thereby can be optimized clustering model,improve the performance of the algorithm.Experimental results on a large number of real data sets show that the proposed algorithm is more effective than other multi-view clustering algorithms.