Research Of Subspace Clustering Algorithm Based On Self-Representation

Cluster analysis has been widely used as an important analysis tool in the field of data mining.With the continuous change of application scenarios,especially entering in the era of "big data",the high dimensionality of data has become more and more common.As high-dimensional data contains a large number of irrelevant attributes,which makes the traditional clustering algorithm based on the original feature space become no longer applicable.Thus the subspace clustering algorithm came into being.In recent years,a subspace clustering algorithm based on self-representation has become a new research hotspot due to its efficient clustering performance.This paper studies the basic principles and algorithms of self-representation subspace clustering,further analyzes the advantages and disadvantages of the existing algorithms,and proposes corresponding improvement methods.The main work of the paper is as follows:Least squares regression model for subspace clustering is one of the classical algorithms for self-representation subspace clustering issues.Since the algorithm treats all data points equally,without considering the correlation information between the data in the solution process,so the representation coefficients between clusters are too dense.To solve this problem,we propose a spatial constraint weighted least squares regression model for subspace clustering.The spatial constraint information is introduced into the objective function of the original least squares subspace clustering algorithm,reconstructing the regular terms by weighted coefficient,so that the representation coefficient weakens the connection between classes while enhancing the connection within classes.The experimental results compared with the state-of-the-art algorithms demonstrate the efficacy of our proposed algorithm.Moreover,most of the current improvements for self-representation subspace clustering are based on coefficient matrix,there is still a lack of research on similarity matrices.Self-representation subspace clustering is an algorithm based on spectral clustering,and similarity matrix is considered as the core of spectral clustering algorithm thus directly affects the result of clustering.This paper focuses on the influence of the construction method of similarity matrix on the performance of clustering.This paper firstly summarizes and compares the current major similarity matrix construction methods,and then proposes improvement methods on the construction of similarity matrix from two different perspectives.First,for the problem of easy bringing noise based on the similarity matrix constructed by the global method,a method based on the local method is proposed.Secondly,as the effect of clustering based on single-structure similarity matrix is not good enough.A similarity matrix construction method based on coefficient matrix fusion is proposed.Finally,experimental results on a large number of datasets show the effectiveness of the improved method in this paper.