基于多核系统的并行线性RankSVM算法

发布时间：2018-07-20 20:07

【摘要】：现有的线性RankSVM已得到较有效的研究,但在训练大规模的线性Rank SVM时,过长的训练时间依然难以让人接受。通过对当前最先进算法Tree-TRON的分析可知,利用信任区域的牛顿迭代(trust region Newton method,TRON)去训练线性Rank SVM模型涉及大量的Hessian-vector内积(Hessian-vector product)计算,同时完成Hessian-vector内积计算又需计算大量的辅助变量和矩阵运算。为了有效地加速与Hessian-vector内积有关的计算,在多核系统下提出了一种高效的并行算法(命名为PRank SVM)用于提高大规模线性Rank SVM的训练速度。PRank SVM的特征主要体现为两个方面:训练数据按不同的查询划分为不同的子问题;在多核系统下,利用多核加速辅助变量和相关矩阵的计算。通过实验分析可知,相较于现有的算法(如Tree-TRON),PRank SVM不仅可以有效地提高训练速度,而且可以有效地确保预测的准确率。
[Abstract]:The existing linear Rank SVM has been studied effectively, but the long training time is still difficult to accept when training large scale linear Rank SVM. Through the analysis of the most advanced algorithm, Tree-TRON, we can see that the Newton iterative (trust region Newton method (TRON) of trust region is used to train the linear Rank SVM model, which involves a large number of Hessian-vector product calculations. At the same time, the calculation of Hessian-vector inner product needs to calculate a large number of auxiliary variables and matrix operations. In order to efficiently speed up the calculation related to the inner product of the Hessian-vector, In this paper, an efficient parallel algorithm named PRank SVM is proposed to improve the training speed of large scale linear Rank SVM. The features of PRank SVM are as follows: the training data is divided into different sub-problems according to different queries; In multi-core systems, the computation of auxiliary variables and correlation matrices is accelerated by using multi-cores. The experimental results show that compared with the existing algorithms (such as Tree-TRON), PRank SVM can not only effectively improve the training speed, but also effectively ensure the accuracy of prediction.
【作者单位】：广东科技学院计算机系;中山大学数据科学与计算机学院;华南农业大学数学与信息学院/软件学院;
【基金】：国家自然科学基金资助项目(61673157) 广东省自然科学基金资助项目(2014A030313454)
【分类号】：TP181

【相似文献】