A Study Of Soft Subspace Cluster Based On Natural Computation

With the coming of the information age,all kinds of availability of high dimensional data are exploded on the Internet.For example,millions of image data are generated from cameras which installed in stations,streets,airports and shopping malls,a large number of stock information data,the gene expression data with high-dimensional,large-scale commodity information data,a great deal of document information data,etc.Those provide the raw materials for the compression,storage,clustering and transmission of complex high-dimensional data.Clustering is one of the key technologies of data mining.The purpose is to study the logical or physical relationships between data,so as to reveal the internal relations and differences between the data,and provide an important basis for further data analysis and knowledge discovery.However,the explosion of high-dimensional data brings a great challenge to the classical clustering algorithms.The defects of those algorithms are more and more obvious.Due to the sparseness and redundancy of high-dimensional data,clusters may exist only in some low-dimensional subspaces.And the similarity measure of the classical methods,such as k-means,is no longer applicable.So it is difficult to predict the result of high-dimensional data with traditional clustering.Cluster high-dimensional data and apply it to practical problems are of great significance in promoting the development of the information age.This paper focuses on the clustering of high-dimensional data.Firstly,it summarizes the existing methods of subspace clustering.Then it makes a research and improvement aiming at the existing deficiencies of subspace clustering methods especially for poor diversity,local optimization and multi-class clustering problem.Finally,this paper put forward three kinds of subspace clustering algorithm for high-dimensional data.The specific content and work arrangements are as follows:(1)In order to improve the accuracy of the algorithm and the ability of global search,so as to avoid the local optimum,this paper proposes a soft subspace clustering based on differential evolution(DESSC)algorithm.The differential evolution algorithm is used to optimize the weight matrix of subspace clustering,which effectively improves the clustering accuracy of high-dimensional data.(2)The traditional cluster algorithm which has the problem of instability and was easily trapping in local optimum was improved.The Quantum-behaved Particle Swarm Optimization(QPSO)algorithm is introduced and combined with subspace clustering algorithm to optimize the weight matrix.And proposes a soft subspace clustering based on QPSO(QPSOSC)algorithm which improves the diversity and stability of weight matrix effectively.(3)Some multi-class high-dimensional data and the number of categories are not always very clear or accurate.This paper proposes a soft subspace clustering based on NSGA-?(NSGASC)algorithm to deal with that condition.The principle of the algorithm is introduced in detail,and the contrast experiments are carried out on the multi-class high-dimensional datasets.