Research Of Parallel Sparse Subspace Clustering Methods Based On Coordinate Descent

Sparse subspace clustering is clustered by finding the similarity between subspace data, and is widely used in computer vision problems such as multi-objective motion segmentation, face clustering and image compression. With the continuous expansion of the scale of data, sparse subspace clustering problem is faced with a huge challenge in computing. The existing sparse subspace clustering algorithm is often based on serial implementation (such as ADMM solution similarity matrix steps), it is difficult to use multi-core processors and computer clusters to improve the efficiency of large-scale clustering problems. Therefore, we proposes a parallel sparse subspace clustering method based on coordinate descent.The concrete work is summarized as follows:Using the coordinate descent method to realize the parallel sparse subspace clustering, an algorithm model of parallel sparse subspace clustering is proposed by using the coordinate descent method. The concurrency characteristic and the coordinate descent method are used to solve the Lasso problem quickly and accurately. Sparse subspace clustering can be modeled to solve the characteristics of a series of sparse self-expression subprograms. Using the coordinate descending method to solve each sub-problem, the parameters are few and the convergence is fast, and the sparse self-expression problem is independent of each other. The experimental results show that the clustering accuracy of the sparse subspace is greatly improved when the clustering accuracy is not reduced, and the parallel sparse subspace clustering is realized by using the openMP parallel framework.The Lasso dynamic screening method based on the infinite norm decision is to use the properties of the infinite norm of the matrix to effectively remove some redundant steps in the iterative updating process. The sparse solution must have a large value of zero when the objective function converges, The method of this section will find out the location of these zero values, skip the optimization process more like the calculation steps in the case of the greater the degree of sparse solution compared to the full update when the speed increase the more obvious, faster to achieve the objective function , The experimental data show that the algorithm proposed in this chapter can achieve very good results in both simulation data and real data.Distributed sparse subspace clustering application based on the spark platform, a sparse subspace clustering system based on the spark platform is designed. The system runs on the linux server cluster of the lab,through a management node and four working node subspace clustering algorithm is distributed and distributed using Spark's MapReduce mechanism. The Spark's distributed computing framework is built on the cluster of computing nodes. The experimental results show that the Spark framework can deal with the problem of large sparse subspace. With the increase of computing nodes, the shorter the time needed to solve the problem, the faster the data can be processed and the solution The unrealability of large-scale data storage and memory allocation in a stand-alone environment.The traditional sparse subspace clustering algorithm is indivisible and the inefficient implementation is proposed. The parallel sparse subspace clustering method based on the coordinate descent is proposed. The key descent method is used to solve the key calculation of sparse subspace clustering. Today's popular parallel and distributed framework to achieve, the experimental results show that the idea of this algorithm using openMP and Spark distributed framework, in dealing with high-dimensional sparse problems can achieve good experimental results.