结合GPU技术的并行CP张量分解算法  

Parallel CP Tensor Decomposition Algorithm Combining with GPU Technology

作　　者：武昱 闫光辉[1] 王雅斐 马青青 刘宇轩 WU Yu;YAN Guang-hui;WANG Ya-fei;MA Qing-qing;LIU Yu-xuan(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730030,China)

机构地区：[1]兰州交通大学电子与信息工程学院,兰州730030

出　　处：《计算机科学》2018年第11期298-303,317,共7页Computer Science

基　　金：国家自然科学基金项目(61662066;61163010)资助

摘　　要：随着高维数据的涌现,张量和张量分解方法在数据分析领域中受到了广泛关注。然而,张量数据的高维度和稀疏特性,导致算法的复杂度较高,阻碍了张量分解算法在实际中的应用。许多学者通过引入并行计算来提升张量分解算法的计算效率。在现有研究的基础上,给出一种简化计算Khatri-Rao乘积的GPU并行CP张量分解算法,称为ParSCP-ALS。在模拟数据集和真实数据集上的实验结果显示,相比现有并行算法,文中设计的ParSCP-ALS算法能有效提高CP张量分解的计算效率,其中在Movielens数据集上的计算时间减少了约58%。With the emergence of high-dimensional data,tensor and tensor decomposition have draw widespread attention in the field of data alanlysis.The high dimensionality and sparsity of tensor data results in a high computational complexity of tensor methods,which becomes an obstacle to the application of tensor decomposition in practical.Many researchers have introduced the parallel computing methods to improve the efficiency of tensor decomposition algorithm.Based on the existing research,this paper presented a GPU parallel CP tensor decomposition algorithm through simply calculating Khatri-Rao product,called ParSCP-ALS algorithm.The experimental results show that the proposed ParSCP-ALS algorithm can effectively improve the computational efficiency of CP tensor decomposition compared with the existing parallel algorithm.On the Movielens data set,the ParSCP-ALS algorithm reduces the computational time by about 58%compared with the existing parallel algorithm.

关 键 词：CP张量分解 CP-ALS算法 GPU CUDA 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]