Application of randomized low-rank approximation algorithm in recommendation system
-
摘要: 针对推荐系统中数据量越来越大,其对应的矩阵填充问题算法效率有待提升. 基于随机算法策略以及高效数据访问要求,提出一种新的求解矩阵填充问题的算法,并借助Matlab软件实现该算法. 数值试验结果表明,该算法在效率上可提升30%左右.Abstract: In view of the problem that with increasing large amount of datas in a recommendation system, the computing efficiency of its corresponding matrix completion algorithms need to be improved. Based on randomized algorithm strategy and efficient data access requirements, a new algorithm for solving matrix completion problem was proposed, and the Matlab software was employed to realize the algorithm. The numerical experiment result shows that the new algorithm can speed up the computing efficiency of original one by about 30%.
-
表 1 矩阵填充问题的比较
Table 1. Comparison of matrix completion problems
数据集 方法 迭代次数 时间/s NMAE 非零奇异值的个数 最大奇异值 最小奇异值 jester-1 IALM-LANSVD 12 6.01 0.184 100 2.14E+03 1 IALM-SVDS 12 3.92 0.182 100 2.15E+03 1 IALM-SPRSVD 12 3.05 0.235 100 2.31E+03 1 jester-2 IALM-LANSVD 12 5.66 0.185 100 2.13E+03 1 IALM-SVDS 12 3.57 0.182 100 2.13E+03 1 IALM-SPRSVD 12 2.67 0.238 100 2.09E+03 1 jester-3 IALM-LANSVD 12 7.50 0.126 99 1.79E+03 1 IALM-SVDS 12 4.19 0.125 100 1.80E+03 1 IALM-SPRSVD 11 3.11 0.157 100 1.88E+03 1 -
[1] 王越, 程昌正. 协同过滤算法在电影推荐中的应用[J] . 兵器装备工程学报,2014(5):86 − 88. [2] 王元涛. Netflix数据集上的协同过滤算法[D]. 北京: 清华大学, 2009. [3] 冯栩, 李可欣, 喻文健, 等. 基于随机奇异值分解的快速矩阵补全算法及其应用[J] . 计算机辅助设计与图形学学报,2017(12):2343 − 2348. [4] FENG Y H, XIAO J W, GU M. Flip-flop spectrum-revealing QR Factorization and its applications on singular value decomposition[J] . Electronic Transactions on Numerical Analysis,2019,51:469 − 494. [5] LARSEN R M. Lanczos bidiagonalization with partial reorthogonalization[J] . DAIMI Report Series,1999,27(537):1 − 101. doi: 10.7146/dpb.v27i537.7070 [6] DRINEAS P, KANNAN R, MAHONEY M W. Fast Monte Carlo algorithms for matrices II: computing a low-rank approximation to a matrix[J] . SIAM Journal on Computing,2006,36(1):158 − 183. doi: 10.1137/S0097539704442696 [7] TROPP J A, YURTSEVER A, UDELL M, et al. Practical sketching algorithms for low-rank matrix approximation[J] . SIAM Journal on Matrix Analysis and Applications,2017,38(4):1454 − 1485. doi: 10.1137/17M1111590 [8] LIN Z C, CHEN M M, MA Y. The augmented lagrange multiplier method for exact recovery of corrupted low-rank matrices[EB/OL]. (2013-10-18)[2021-01-18]. http://arxiv.org/pdf/1009.5055v3.pdf.