Research On Clustering Models And Algorithms Of Highly Dimensional Complex Data

With the advent of the information age,all kinds of data are filled in people's lives.Especially high-dimensional data are ubiquitous in many areas of machine learning,signal and image processing,computer vision,pattern recognition,etc.However,the increasing of dimensions of the data not only increases the computational time and memory requirements,but also arouses the “curse of dimensionality” easily.Especially when the high-dimensioanl data are complex with data nuisances,the traditional clustering algorithms have already been difficult to analyze data.For the problem,the paper researches the clustering algorithms based on the swarm intelligence algorithms and subspace,respectively,and the main contents are listed as follows(1)For flaw of the speed of clustering,we proposed a novel genetic and particle swarm clustering algorithm based a pattern reduction.For fully combining the pattern reduction method,a generalized genetic algorithm be used to improve the particle swarm optimization algorithm.This ways not only can increase diversity of samples,but also can protect these patterns which needs to save in the phase of compression operation.At the same time,a incremental strategy for the particle's numbers is used to replace the ‘poor' particles,which could fully embody the ability of rapid search optimization of the particle swarm algorithm and the advantage of the large search space of the genetic algorithm.The experimental results show that this algorithm could decrease the time of clustering effectively without obvious accuracy decline.(2)For the problem of none of the modified sparse subspace clustering could meeting the property of sparseness between clusters and consistency within cluster perfectly,an evolving iterative weighting(reweighted)l1 minimization framework be proposed,which contains the characteristic of arctan and logarithmic function at the same time.The evolving reweighted l1 minimization framework could simultaneously satisfy the two main features of the l0 minimization framework,which makes a better approximation than original reweighted l1 minimization.Following the evolving reweighted l1 minimization framework,we propose a evolving reweighted sparse subspace clustering.The experiments show that the proposed algorithm could achieve the better performance than other subspace clustering algorithms.(3)In view of the existing sparse subspace clustering algorithms failed to consider the data prior information,this paper establishes a semi-supervised clustering framework on the sparse subspace algorithm and a sparse subspace clustering algorithm based semi-supervised learning is proposed.Firstly,a constraint matrix which is suitable for sparse optimization program is established by a small amount of prior information.Then,according the differentstatuses of representation matrix,the appropriate bound terms be designed to built the sparse optimization program,which could be ensured to acquired the the sparse representation matrix under the guidance of constraint information.Finally,the final clustering results are acquired by spectral clustering.The experiments show that the proposed algorithm not only can hold the clustering accuracy of original algorithm without constraint information,but also increases gradually the clustering accuracy with the augment of constraint information.Compare to the existing semi-supervised clustering algorithms,the proposed algorithm can get higher clustering accuracy.