Research On Improved Sparse Subspace Clustering Algorithm

Cluster analysis aims at partitioning data objects into several meaningful subsets,called clusters, so that data objects are similar to those which are in the same cluster and dissimilar to those which are in different clusters. Clustering high-dimensional data has been a challenging problem in data mining and machine learning.There are two main problems in high-dimensional data clustering: ⑴..those irrelevant properties appeared in high-dimensional data make clustering tendency not obvious. ⑵. classification boundaries in high-dimensional data are not obvious. Aiming to these difficulties, this paper has spent plenty of time on high-dimensional data clustering based on the traditional clustering algorithms, especially in sparse subspace clustering algorithm, in which Sparse Subspace Clustering(SSC) algorithm is outstanding.Sparse Subspace Clustering(SSC) algorithm is superior to other methods in clustering high-dimensional data. And it can deal with noise, missing values and sparse outlying entries in the data sets. For the issue of weight matrix construction in high-dimensional data clustering, we consider the complete information of the solution coefficient vectors of two objects to analyze the similarity between these two objects instead of directly using a particular single sparse coefficient, which only considers local information. Besides we apply the similarity into a new Spectral Clustering algorithm instead of the Spectral Clustering algorithm in SSC. At last, Evaluation experiments on real-world data sets in motion segmentation show that our modified algorithm based on SSC called Modified Sparse Subspace Clustering(MSSC) which use our novel weight matrix construction outperforms the traditional algorithms.