QPR-NN:一种结合二次多项式回归与神经网络的推荐算法  

QPR-NN:A New Recommendation Algorithm Combining Quadric Polynomial Regression and Neural Network

作　　者：廖彬 张陶[3,4] 于炯[3] 国冰磊 李敏[2] 刘炎[5] LIAO Bin;ZHANG Tao;YU Jiong;GUO Binglei;LI Min;LIU Yan(Institute of Silk Road Economy and Management,Xinjiang University of Finance and Economics,Urumqi 830012,China;College of Statistics and Data Science,Xinjiang University of Finance and Economics,Urumqi 830012,China;School of Information Science and Engineering,Xinjiang University,Urumqi 830046,China;Department of Medical Engineering and Technology,Xinjiang Medical University,Urumqi 830011,China;School of Software,Tsinghua University,Beijing 100084,China)

机构地区：[1]新疆财经大学丝路经济与管理研究院,乌鲁木齐830012 [2]新疆财经大学统计与数据科学学院,乌鲁木齐830012 [3]新疆大学信息科学与工程学院,乌鲁木齐830012 [4]新疆医科大学医学工程技术学院,乌鲁木齐830012 [5]清华大学软件学院,北京100084

出　　处：《西安交通大学学报》2019年第9期79-87,136,共10页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(61562078,61462079);新疆维吾尔自治区自然科学基金资助项目(2016D01B014)

摘　　要：针对传统推荐算法不能很好地适应数据高规模及高稀疏性的问题,结合深度学习数据建模的方法,提出了一种结合二次多项式回归与神经网络(QPR-NN)的推荐算法。在对已有特征提取方法缺陷分析的基础上,利用二次多项式回归模型将用户对物品的评分数据进行特征提取及降维,充分挖掘了用户与物品之间的相关性。将特征提取后的数据作为深度学习训练模型的输入,增加输入数据与训练模型之间的匹配度,并将训练得到的模型用于推荐评分预测。在MovieLens与Epinions两组数据集上的实验结果表明:QPR特征提取方法与QPR-NN推荐算法在平分绝对误差与均方根误差评价指标上均优于现有的主流算法,QPR-NN推荐算法可以有效提升推荐准确率。The traditional recommendation algorithms cannot adapt to the problem of high-scale and sparse data. A recommendation algorithm combining quadric polynomial regression and neural network(QPR-NN) is proposed in view of the generality of deep learning data modeling, where quadratic regression is combined with neural network. Following the analysis on the defects in the existing feature extraction methods, the algorithm chooses quadratic regression model to extract feature and reduces the dimensions for the user rating data to fully explore the correlation between the user and item data. The data after feature extraction are taken as the input of deep learning training model to increase the matching degree between the input data and the training model, then this model is used for recommending score prediction. For the two datasets of MovieLens and Epinions, the experimental results show that the QPR feature extraction method and the QPR-NN recommendation algorithm are superior to the existing mainstream algorithms in the evaluation indexes mean absolute error and root mean square error, and effectively improve the recommendation accuracy.

关 键 词：推荐算法 深度学习 特征提取 二次多项式回归 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]