加速交替最小二乘法推荐系统优化设计  

The Optimal Design of an Accelerated Alternating Least Square Recommendation System

作　　者：张宏烈[1] 刘佳星 刘艳菊[1] 张惠玉 ZHANG Hong-lie;LIU Jia-xing;LIU Yan-ju;ZHANG Hui-yu(Computer and Control Engineering College,Qiqihar University, Qiqihar 161006, China;Communication and Electronic Engineering College,Qiqihar University, Qiqihar 161006, China)

机构地区：[1]齐齐哈尔大学计算机与控制工程学院,齐齐哈尔161006 [2]齐齐哈尔大学通信与电子工程院,齐齐哈尔161006

出　　处：《科学技术与工程》2019年第14期257-261,共5页Science Technology and Engineering

基　　金：黑龙江省教育厅基本业务专项理工面上项目(135209234);黑龙江省自然科学基金面上项目(F201440)资助

摘　　要：推荐系统帮助用户在海量数据中更便捷地找到他们最感兴趣的内容。但推荐系统存在可信度低、推荐结果的可解释性不足、可扩展性不好、随着用户数量的增大,计算时间增长且精度较低、数据稀疏性和冷启动等问题。为此提出基于交替最小二乘法(alternating least squares,ALS)的推荐系统优化算法,在ALS基础上对两个部分进一步优化:第一部分采用LBFGS (limited-memory broyden-fletcher-goldfarb-shanno)算法使搜索方向快速计算出来;第二部分采用阻尼牛顿法求解步长因子。在Spark平台上加以验证,取得较好效果。The recommendation system has emerged in the era of big data,and its core meaning is to help users find the content that they are most concerned from the massive data. Although the recommendation system has a wide range of application scenarios,it has the defects of low credibility,insufficient recommendation result interpretability,poor scalability,user number increase,computation time increase,low computation precision,data sparsity and cold start. A recommendation system optimization algorithm based on alternating least squares( ALS)is proposed,which further optimizes the two parts based on ALS. Limited-memory broyden-fletcher-goldfarb-shanno( L-BFGS) algorithm was adopted to calculate the search direction quickly,then the damped Newton method was used to solve the step factor. The proposed algorithm is verified on the Spark platform and the favourable results are obtained.

关 键 词：推荐系统 交替最小二乘法 L-BFGS 阻尼牛顿法 SPARK 

分 类 号：TP391.45[自动化与计算机技术—计算机应用技术]