| Since the social science and technology rapidly updating,the value of data has become more precious.As a result,all walks of life are starting or planning to store large amounts of data.In order to make the data reflect value,many techniques of data mining and data analysis have become the research hot spots in recent years.And the algorithms of classification and clustering are the two main branches.Clustering is to divide similar samples into the same category without labels.Up until now,many clustering methods of various working mechanisms have been proposed,among which graph-based algorithms play an important role.The graph-based approaches utilize the nonlinear pair similarity between data points for clustering,such as spectral clustering(SC).Benefited from the construction of similarity matrix and the cutting of graph,this algorithm can divide arbitrary data samples and obtain better clustering accuracy.However,there are some drawbacks to this type of approach.First of all,although SC has the property of dimensionality reduction to some extent and has a good effect on some high-dimensional data,the effect may not be so obvious when the data dimension is too high.Secondly,similar matrix is unreasonable in many cases because it takes a lot of time to construct the matrix and it is difficult to effectively process large-scale data.As the great improvement of remote sensing imaging technology,the pixel of hyperspectral image(HSI)is gradually increased,and it appears more and more frequently in the actual situation monitoring on the ground,precision agriculture,military field and other scenes.With the rapid increase of the amount of HSI data,reducing the computational complexity of clustering while improving the accuracy of HSI clustering has become one of the research hotspots.In order to better apply the clustering technology to hyperspectral image,a new joint clustering model,fast spectrum embedded clustering based on structured graph learning,is proposed.Firstly,the embedded low-dimensional data is initialized by singular value decomposition of the similarity matrix of the anchor graph.Then,the low dimensional representation method is used to solve the similarity matrix,and structured graph learning is used to update the similarity matrix.Secondly,the anchor graph matrix is modified to update the embedded data through the external connection method,so that the algorithm has better performance in the hyperspectral image data set.In order to better apply the clustering technology to high-dimensional data,this paper also proposes a clustering model combining bipartite graph and structured graph learning,consisting of two main processes.Based on bipartite graph embedding,the low-dimensional representation of data points can be obtained.In addition,through structured graph learning,the optimized similarity matrix and the clustering results can be achieved at the same time.By using the similarity matrix,the structure of bipartite graph can be modified by iteration,so as to obtain better low-dimensional representation.The results show that the proposed method has good performance in high dimensional data,and it saves time compared with the traditional spectral based methods in large scale data.In addition,this method learns the similarity graph and gets the clustering results simultaneously,which overcomes the optimization limitation of two separate processes in the traditional spectral clustering. |
