一种面向矩阵分解模型的推荐系统训练加速方法  

Algorithmic acceleration of matrix factorization based recommendation system

作　　者：段圣宇 吴伊宁 赛高乐 DUAN Shengyu;WU Yining;SAI Gaole(School of Computer Engineering and Science,Shanghai University,Shanghai 200444,China;College of Integrated Circuits and Optoelectronic Chips,Shenzhen Technology University,Shenzhen 518118,Guangdong,China)

机构地区：[1]上海大学计算机工程与科学学院,上海200444 [2]深圳技术大学集成电路与光电芯片学院,广东深圳518118

出　　处：《山东大学学报(工学版)》2025年第1期24-29,共6页Journal of Shandong University(Engineering Science)

基　　金：计算机体系结构国家重点实验室开放课题资助项目(CARCH201909)。

摘　　要：为降低矩阵分解(matrix factorization,MF)模型面向推荐系统应用的训练时间,特别针对细粒度稀疏的特征矩阵在训练过程中存在大量无效乘法运算的问题,提出一种基于特征矩阵联合稀疏性进行近似计算的训练加速方法。基于隐因子向量稀疏性强弱基本不变的特点,提出在模型训练初期,根据隐因子向量的稀疏性,对特征矩阵重新排列;在训练过程中,采用早停法,避免无效乘法运算。试验结果表明,模型训练过程中乘法运算次数可最多降低28.41%,加速前后评分预测值相关系数约0.95。所提出方法可以保证预测准确性小幅降低的同时,显著减少训练中的乘法运算次数,针对更大规模的矩阵分解模型训练,能实现更好的加速效果。In order to reduce the training time of matrix factorization(MF)based recommendation system,specifically considering the fine-grained structured sparsity of the decomposed matrices,which caused unnecessary multiplications and increased the overall time of training process,an algorithmic acceleration method,based on joint sparsity of the decomposed matrices and approximate matrix multiplications,was proposed.According to an observation that the trends of sparsity on all latent vectors generally hold,an algorithm to rearrange the feature matrices during the first a few training epochs,based on joint sparsity,was proposed.An early stop algorithm was applied to eliminate unnecessary multiplications during the training process.The experimental results showed the total number of multiplications could be reduced by up to 28.41%,and the correlation between the predicted ratings produced by the conventional and proposed methods was around 0.95.The acceleration method could greatly reduce the total number of multiplications during MF training process,causing a minimal error,and more multiplications were expected to be eliminated for the recommendation systems with larger scales.

关 键 词：推荐系统 矩阵分解 稀疏性 算法加速 近似计算 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]