The Application Of Spectral Clustering With Neighborhood Information

Cluster analysis has a very wide range of applications in information retrieval anddata mining and so on, which is an effective strategy to understand and explore theintrinsic link between things. Clustering algorithm can divide the data into differentclusters without any prior knowledge and find out the valuable information from thedata. Spectral clustering, as a new and high-performance clustering method, which hasbeen widely used in computer vision, speech recognition, text mining and other fields.Spectral clustering algorithm based on spectral graph theory is to transform theclustering problem into a graph optimal partitioning problem, and it is a pointwiseclustering problem, so it has good application prospects for data clustering. Comparedwith the traditional clustering algorithms, such as K-means and EM algorithm and so on,spectral clustering algorithm can cluster datasets in any shape sample space andconverge to the global optimal solution, and also avoid the singularity problem causedby the high dimensionality of data.Spectral clustering algorithm is a clustering algorithm based on similarity matrix,which utilizes the spectral graph theory to divide the similarity matrix. Firstly, thesimilarity matrix W is constructed by the similarity measure of the sample dataset intraditional spectral clustering algorithm. Secondly, we can obtain the Laplacian matrix Lbased on matrix W, and then calculate the eigenvalues and eigenvectors of the matrix L.Finally, we can select one or more feature vectors to cluster different data. The spectralclustering algorithm is directly to cluster the similarity matrix, so different forms ofsimilarity matrix will have great impact on the algorithm performance. Therefore, howto construct a suitable similarity matrix has become an important research questions inspectral clustering algorithm.This paper introduces the basic principles of spectral clustering, as well astraditional spectral clustering algorithms and classification. Meantime, we analyzeexisting problems and challenges about spectral clustering. Then, how to take advantageof neighborhood information as a starting point, two different methods are proposed forour subject. The achievements are as follows.1. A fuzzy spectral clustering algorithm based on neighborhood information isproposed. The method utilizes fuzzy local information C-Means(FLICM) clusteringalgorithm to cluster eigenvectors in the last step of spectral clustering algorithm. FLICMalgorithm based on Fuzzy C-Means (FCM) algorithm, which overcomes the effects of noise points by adding and making full use of the neighborhood information of samplepoints. Furthermore, FLICM algorithm can improve the performance of anti-noise ofspectral clustering algorithm by utilizing neighborhood information. Then test results onUCI data sets and change detection of remote sensing images demonstrate the methodwe proposed is feasible and effective.2. A spectral clustering algorithm based on bilateral integration is proposed. As weall know, the structure of the similarity matrix plays a very important role in spectralclustering algorithm, so its quality directly affects the results of clustering. Consideringit, in this method, firstly, subtraction-similar matrix and neighborhoodsubtraction-similarity matrix are constructed.Secondly, we make use of the idea ofbilateral fusion to fuse the two similar matrix, fusion similarity matrix obtained takesfull advantage of neighborhood information of pixels and gray information. Therefore, itis better to suppress the effect of noise points. We apply this method for remote sensingimage change detection, and experimental results demonstrate the improved spectralclustering algorithm greatly outperforms the traditional spectral clustering algorithm interms of accuracy.