| Support Vector Regression Machine (SVR) is the application of Support VectorMachine (SVM) in regression problem and shows good generalization andpromotion ability. In modern learning theory, time efficiency and the accuracy ofresults are always the goals to pursue. With the coming of big data era, inconsideration of the algorithm time complexity of SVM, when serial SVM meets thelarge scale training samples, it would take long time to training SVM task.In order to improve the training efficiency and keeping the accuracy of results, Ihave analyzed and concluded the theory of SVM and the development history ofDistributed Parallel Support Vector Machine (DPSVM), then propose a BSP-BasedDistributed Parallel Support Vector Regression Machine (DPSVR) Framework.SVR has a strong robust regression when bringing inεinsensitive loss function,and it makes SVR to own the classification accuracy rate and detection rate. So theSVR can be used in many fields. Due to SVM problem is a stochastic problem, thereis no perfect DPSVR algorithm to solve the general problem. After analyzing thedevelopment history of DPSVM, this paper obtained that the main difference ofDPSVM algorithm is the topology among distributed computing nodes. In order toreduce the computing resource and improve the expansibility of framework, thepaper has optimized the framework by adding dynamic changing topology graph.Then, a self-adaptation cascade DPSVR algorithm has been designed andimplemented to verify the expansibility of the framework and improve the generalefficiency of cascade DPSVR algorithm.In the paper, four classical DPSVR algorithms have been chosen to verify theavailability and performance of the framework. The experiment data derived fromKDD99dataset in intrusion detection domain, and the experiments are performed onhigh-performance computer of ZQ3000cluster. The results on this paper proved theavailability and performance of framework and the self-adaptation cascade DPSVRalgorithm. |
PDF Full Text Request