Research On Dimensionality Reduction Methods In Edge Computing

With the rapid development of the Internet of things,a large number of data are generated,which brings huge data processing pressure to cloud computing.As an extension of cloud computing,the abilities of data processing and storage are extended to the edge side of the network near the Internet of things devices by edge computing.With the edge computing,a large amount of data don't need to be uploaded to the cloud,which can reduce the transmission delay.However,there are two urgent problems need to be solved: one is how to achieve task processing and feedback in a timely manner near the data source.The other is that how to alleviate the high dimension of data when the distribution of samples is uneven in each dimension space.The service quality of edge computing is directly determined by these problems and they also determine users' satisfaction.Therefore,the dimension reduction method in edge computing is chosen as the research topic.Considering the edge computing architecture,the edge environment,the effect of dimension reduction,execution time and so on,two dimension reduction methods are proposed.The purpose is to remove redundant information,extract the main features of data,save users' s analysis time and make reasonable judgments.The specific research contents are as follows.(1)The dimension reduction algorithm of edge computing based on adaptive optimization of neighborhood set selection is designed.In order to solve the problems of easy deformation of the manifold structure,poor embedding effect and poor user experience in the process of dimension reduction.Firstly,a four-tier data processing architecture in the edge environment is designed to perform edge computing tasks.And then,a multi-group weight local linear embedding algorithm based on adaptive optimization neighborhood set selection is proposed.The upper and lower thresholds are set to filter the data.According to the curvature of the manifold and the sample density,the neighborhood value is dynamically selected.Finally,a single weight is replaced by multiple sets of linearly independent weights to construct the local structure.Through the experimental evaluation on three groups of classic data sets,it is proved that this method has a better dimension reduction effect and shorter task processing time than the LLE method.(2)The research of multi-step group dimension reduction framework in edge computing.The framework consists of two parts: LTRBM-SGA and TAA.Firstly,for the characteristics of“large attributes and small samples” of some edge environment data,a multi-step grouping dimension reduction algorithm based on loss threshold and restricted Boltzmann machine is designed.By adopting the “multi step and group” method,tasks are grouped and distributed to the edge server according to the number of attributes.For the tasks received by the edge server,covariance matrix is used to measure the relevance and heterogeneity between each group of data attributes and information loss threshold function that is set to remove redundant dimensions in big data.The restricted Boltzmann machine is used for feature extraction,modeling and classification.Secondly,the task allocation algorithm is designed to improve the task completion rate and enhance the stability of the edge environment.Finally,experiments show that the task allocation algorithm can effectively improve the task completion rate.Compared with RPCA and CNN,the multi-step group dimension reduction algorithm based on loss threshold and restricted Boltzmann machine is more accurate and the task completion time is effectively reduced.