一种大数据集上的非线性PSVM训练方法

PSVM作为一种新型SVM方法，避免了求解二次规划问题，具有更快的计算速度，但对于大规模数据集，采用传统方法求解非线性PSVM面临大矩阵求逆的困难。文章基于共轭梯度法结合低秩估计提出了一个大数据集上的非线性PSVM训练方法NPSVM-LD，通过多次迭代的矩阵乘积运算避免了对大矩阵的求逆。在UCI数据集上的实验表明。该方法能够在应用非线性核函数条件下，使PSVM有效处理规模在10000以内的训练集的情况。

A Training Method for Nonlinear PSVM on Large Datasets

As a new method of SVM, PSVM works faster by avoiding solving quadratic programming problems. However, to solve nonlinear PSVM by traditional method has the difficulty in inverting a large-scale matrix. In this paper we present a training method for nonlinear PSVM on large datasets-NPSVM-LD which is based on the conjugate gradient method combined with low rank approximation. The method avoid inverting a large-scale matrix by iterative matrix multiplications. Experiments on UCI dataset indicate that the method enable nonlinear PSVM to tackle training sets whose scale is no more than 10000 efficiently.