Improved Support Vector Machine Classify Method Based On Transductive Learning And Transfer Learning And Its Application

Support Vector Machine (SVM) is an effective classification method based on thesmall sample-size learning theory，and it has become a hot topic of the machine learningcommunity. In my paper, I researches Transductive Support Vector Machine (TSVM)and Support Vector Machine based on transfer learning, and improves this two methods.In this paper, we first expatiate the current research of Radar Emitter classifymethod and SVM method, and then introduce basic concept of SVM and transferlearning. Then TSVM and Processing Transductive Support Vector Machine (PTSVM)is studied, and in allusion to the shortcoming of long training time of PTSVM, we addthe idea of K nearest neighbor and cache to PTSVM, deal with multiple samples oneach iteration process of this algorithm, which reduces training time of the algorithm alot.KMMSVM algorithm first uses Kernel Mean Match (KMM) algorithm to calculateeach source samples’ weight value, then chooses some source samples whose weightvalue are greater than the threshold, and uses chosen source samples to train SVMclassifier and classify the target samples at last. When the number of source samples islarge, KMMSVM algorithm takes too much time. In allusion to this shortcoming, thispaper integrates ensemble leaning with KMMSVM algorithm, constructs multiple baseclassifiers, then ensembles results of single base classifiers in a certain way. Each baseclassifier’s train samples only contain a small amount of source samples, so the trainingtime of the algorithm reduces a lot, and the ensemble of base classifiers’ resultsimproves the classification accuracy at the same time.Finally, the improved PTSVM and KMMSVM algorithms are applied to the RadarEmitter Recognition. Through feature extraction on Radar Emitter data, thePretreatment radar simulation data is used as experimental data sets, and then comparesthe improved algorithms and the original algorithms by using the radar data sets,experiment shows that the improved algorithm is better than the unimproved algorithm.