支持向量回归机训练集的并行预处理方法

为加快支持向量回归机在求解大样本集问题时的训练速度，提出了并行支持向量回归机。该方法根据核矩阵把数据集分成k个子集，通过并行预处理过滤掉非支持向量，再对剩余的支持向量进行训练得到决策函数。实验表明，本算法不仅预测准确度跟标准的分解算法基本一致，而且大大缩减训练时间，具有很高的加速比，同时需要的训练时间大大少于Graf等人提出的级联结构的算法，另外，算法还可有效地缩减支持向量的数目。

Parallel Pretreatment Method for Training Set of Support Vector Regression Machine

For accelerating the training speed of support vector regression （SVR） machine in solving large scale problems, a novel parallel SVR machine was proposed. The parallel SVR machine first partitioned the whole training data set into k smaller subsets according to the kernel matrix, then, filtered the non-support vectors by using parallel pretreatment method and finally, trained the remained support vectors to acquire decision function. Experiments show that this algorithm not only is comparable to standard decomposition algorithm in terms of predicting precision, but also greatly reduces training time, and it is of very high rate of speeding up and needs less time than that of cascade architecture algorithm proposed by Graf et al. And this algorithm also effectively reduces the number of total support vectors.