A Spectral Clustering Based on Locally Linear Embedding

Background: With the rapid development of information technologies, digging out useful information from mass data has become a hot issue. We should cluster the data before the analysis. Human clustering of mass data cannot meet the requirement of data mining, therefore, various auto clustering algorithms come out successively. Spectral Clustering is a commonly-used cluster algorithm and the effect of spectral clustering highly depends on similarity matrix. Gaussian kernel method has the problem with selecting the good parameter. In real world data set, there is always noise. It is hard to select a good parameter to construct an ideal similarity matrix by Gaussian kernel function. Method: This paper proposes a similarity matrix constructing method based on locally linear embedding. This kind of graph is sparser than Gaussian method and has little noise. This method is not sensitive to noise compared with Gaussian kernel function. The experiments on real world data sets prove the effect of this method. Result: This paper starts from the locally linear expression relationship, uses the non-negative linear value constructing similarity matrix and gets a better experiment result.