基于径向基函数神经网络预抽取的支持向量机

SVM based on Radial Basis Function Neural Networks Pre-extraction

标准的支持向量机训练算法实际上是一个带约束的QP问题,在QP问题求解时,其存储需求随着训练样本数量的平方(甚至立方)增长,而在实际的应用中,可能涉及数千个甚至更多数据样本,受存储容量的限制,支持向量机将很难直接处理这样的大规模数据任务.径向基函数神经网络具有良好的收敛性及快速训练的优势,为了减少大规模数据集下支持向量机的训练时间,论文提出一种基于径向基函数神经网络预抽取的支持向量机RBFPE-SVM.在径向基函数神经网络的设计中,采用自适应k-均值聚类及递归最小二乘算法计算权向量的混合结构,可以获得最优的决策边界,RBFPE-SVM通过径向基函数神经网络决策边界从原始数据集中抽取支持向量候选集,该候选集规模远小于原始训练数据集,研究结果表明,RBFPE-SVM与标准支持向量机S-SVM相比,能达到较低的训练时间和内存要求,支持向量少,保证了其泛化能力和分类速度.

The standard SVM training algorithm leads to a quadratic optimization problem(i.e.,QP problem)with bound constraints and one linear equality constraint.However,the memory requirements of the QP problem grow with the square(or even cube)of the number of training samples.Consequently,in real-life applications,which may involve several thousand data samples,direct attempts to solve the QP problem in an SVM will not scale to the large task.Because the RBFNN has good convergence and fast training,to reduce the training time of SVM for a large dataset,a support vectors(SV)pre-extraction method based on a radial basis function neural network(RBFNN)is proposed in this paper.In the present design of the RBFNN,the adaptive k-means algorithm for clustering computation is applied,followed by recursive least squares(RLS)for computing the weight vector.Thereafter,a candidate set can be extracted from the original training set using the decision boundary.It is worth noting that the candidate set is far smaller than the original training set.The experimental results show that the proposed method improves the performance of the SVM,which ensures its generalization ability and classification speed.