基于层次K-均值聚类的支持向量机模型

摘要

This paper presents an improved SVM learning model,it is based on hierarchical k-means clustering and is called as hierarchical k-means SVM (HKSVM),to solve the problem of SVM in low classification efficiency.The method first makes k-means clustering on every sample class respectively and calculates the centre of each class as well as trains SVM to get initial classifier;then it divides the clustering result into active class set and static class set according to the relationship between the hyperplane and the clustering result,and conducts deeper clustering on the active sets near to the hyperplane for obtaining even smaller classes,and calculates at the same time the class centres to train new SVM model,and corrects the classified hyperplane.This process is on ad infinitum until the more precise classifier is obtained.Adopting hierarchical k-means clustering-based SVM model and by incessant deep clustering on active class sets,more sample points are obtained near the classified hyperplane;however,in where farther from the hyperplane,the extracted training samples are not much so as to effectively compress the size of the training set,and significantly improve SVM’s learning efficiency while keeping its training precision.The experimental results on UCI benchmark datasets demonstrate that the proposed HKSVM model achieves higher classification efficiency and testing accuracy simultaneously on large-scale dataset.%针对支持向量机SVM分类效率低下的问题，提出一种基于层次K-均值聚类的支持向量机HKSVM（Hierarchical K-means SVM）学习模型。该方法首先对每类样本分别进行K-均值聚类，计算每类中心并训练SVM，得到初始分类器；然后根据超平面与聚类结果的关系，将聚类所得结果划分为活动类集和静止类集，并对超平面附近的活动类集进行深层聚类，以得到更小的类别同时计算类中心来训练新的SVM模型，并校正分类超平面，如此循环往复，直到得到较为精确的分类器为止。采用基于层次K-均值聚类的SVM模型，通过对活动类集进行不断地深层次聚类，从而在分类超平面附近得到较多样本点，而在距离超平面较远处则取少量训练样本，以有效压缩训练集规模，在保持SVM训练精度的同时大幅度提高其学习效率。标准数据集上的实验结果表明，HKSVM方法在大规模数据集上同时得到了较高的分类效率和测试精度。