Research On Local Kernel Classifiers And Its Application Of Pulse Classification

Local Learning algorithm has a small generalization error and is conerned inMachine Learning filed in recent years. Compared with the global classifier, It paysattition to the local distribution of samples and the better accuracy can be obtained by aproper choice of the locality parameter in the local classifier, which makes locallearning algorithms very appealing for pattern recognition. A novel Local Kernelmethod (FaLK-SVM) will be introduced and researched in detail and it is a effectiveLocal Kernel Classifier method. In training phase, it uses cover tree to get k'-nearestneighbor and uses a greedy algorithm to get the center covering set to train the modelcentered at this centers with every correspnding k nearest neighbor. In testing phase,given a unknown sample, it uses cover tree to get nearest neighbor and use the localmodelto classify the unknown sample. But it is difficult to select a proper local classifierand because it use greedy algorithm to get k'nearest neighbors, the nearest trainingsample off the unknown sample may be in several local classifiers. Therefore, wepropose a Adaptive Weighted Fusion method of Local Kernel Classifer (FaLK-SVMa)which use the several local models to predict unknown sample. Because of aboveimprovance, it may be make it more stable. In addition, we propose two strategies forweigh and make it more proper to use a distribution of each local model in which thenearest neighbor off the unknown sample.We apply the improved Local Kernel approach named Adaptive Weighted Fusionof Local Kernel Classifiers on UCI data sets of two classes and multiclasses. Twoclasses problem contains forteen small data sets and three large data sets andmulticlasses problem has three multiclasses data sets. It can be seen from experimentalresults, the novel method has obvious advantages not only in higher classificationaccuracies than Local Kernel method in performance but in no icrease time complexity.Therefore, the obvious advantage can be seen from FaLK-SVMa.In order to apply above method on the study of pulse classification, We studytraditional medicine pulse waveform classification based on Local Kernel Classifiers.We use the pulse data from Harbin 211 hospital and make series processes, for instance,preprocessing, extracting single pulse waveform and so on. Then We applyFaLK-SVMa on the pulse waveform data sets and make a lot of experiments to compareLib-SVM, FaLK-SVM with FaLK-SVMa, which is a novel method we proposed namedAdaptive Weighted Fusion of Local Kernel Classifiers. Moreover, We study the effect oflocal parameters on these four methods. It can been seen from experimental results onpulse classification that FaLK-SVMad and FaLK-SVMar can not only get betteraccuracies in two classes pulse waveform classification, but all reach 92.43% in multiclasses problem. In addition, we compare FaLK-SVMad and FaLK-SVMar withother classification methods and It can be seen from results based on 3-fold crossvalidation, the accuracies of FaLK-SVMad and FaLK-SVMar are higher than others andreach 92.27% and 92.23% respectively.