基于多核系统的并行线性RankSVM算法  

Efficient parallel algorithm for linear RankSVM on multi-core systems

作　　者：聂慧[1] 彭娇[1] 金晶[2] 李康顺[3] 

机构地区：[1]广东科技学院计算机系,广东东莞523000 [2]中山大学数据科学与计算机学院,广州510006 [3]华南农业大学数学与信息学院/软件学院,广州510006

出　　处：《计算机应用研究》2017年第1期46-51,57,共7页Application Research of Computers

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)

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。Many effective linear RankSVM algorithms have been studied extensively. However, if making use of any one of them to deal with the large-scale linear RankSVM ,then it must be taken extremely lengthy training time. According to the analysis of the existing state-of-the-art algorithm Tree-TRON, if used trust region Newton method （TRON） to train the linear RankSVM, massive Hessian-vector products and the computation of the auxiliary variables could affect the training speed significantly. To efficiently accelerate these computations, this paper proposed an efficient parallel algorithm （ named PRankSVM） on multi-core systems. All in all, two important issues should be well handled when designing PRankSVM on multi-core systems. First, it divided the training set into several subsets in terms of different queries. Second, it efficiently utilized the great computational power of the multi-core system to improve the Hessian-vector products and the computation of the auxiliary variables. The experimental results show that PRankSVM not only can obtain the excellent convergence speed, but also can ensure the ac- curacy in prediction, while comparing with the existing methods.

关 键 词：排序学习 线性RankSVM模型 并行计算 多核系统 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP301.6[自动化与计算机技术—控制科学与工程]