Improved Spectral Clustering Algorithm And Its Application Research

In the era of big data,data mining has become one of the most active areas.In this field,clustering analysis which can classify data set to be clustered into different classes according to the similarity between data points plays an important role.The spectral clustering algorithm is based on the theory of spectral graph Division.With the help of its solid theoretical basis,convenience for people to understand the clustering principle,easy implementation and wide application scenarios,the spectral clustering algorithm has attracted much attention of researchers.After studying the classical spectral clustering algorithm and related theories,the spectral clustering algorithm is further improved.The main works are as follows:1)This section summarizes the related knowledge of spectral graph theory which the spectral clustering algorithm based on,the research status of spectral clustering algorithm is mainly analyzed.Finally,the improvement direction is put forward in view of the research hot spots.2)A spectral clustering algorithm based on adaptive similarity matrix is proposed.In order to solve the influence on the clustering results,which is caused by the fluctuation of the scale parameter in gaussian kernel function,two concepts of geodesic distance and density are introduced,and a non-parameter adaptive similarity matrix is constructed,which is used to express the similarity between data points in clustering data sets.Experimental verification is carried out on 5 artificial datasets and 5 standard UCI datasets.Under the same condition,compared with the traditional NJW algorithm and K-means algorithm,the results show that the improved algorithm has the highest accuracy and the advantage of clustering for complex structural data.3)When the spectral clustering algorithm is applied to the image segmentation,it needs to consider all the pixels of the image,resulting in a large amount of computation of the similar matrix,and sometimes even overflowing in memory can not continue to run.The Nystrom method in the integral equation solves this problem well,which collects a certain number of sample points randomly,searches for the similarity between sample points and the similarity between sample points and non sample points,and then uses these two similar relationships to approximate the similarity of all pixels.When constructing the similarity matrix of the above two kinds of similarity relations,distance measurement adopts cosine function.In order to overcome the dependence of the K-means algorithm on the initial value,the propagation neighbor clustering algorithm(AP)is used to cluster the obtained low dimensional quantum space.Finally,the superiority of the algorithm is verified in real images.