SVD++推荐算法的超参数  

Super Parameters of SVD++ Recommendation Algorithm

作　　者：张振朋 王以松[1] 冯仁艳 李倩倩[1] ZHANG Zhenpeng;WANG Yisong;FENG Renyan;LI Qianqian(College of Computer Science and Technology,Guizhou University,Guiyang 550025,China)

机构地区：[1]贵州大学计算机科学与技术学院,贵州贵阳550025

出　　处：《贵州大学学报（自然科学版）》2018年第3期97-100,共4页Journal of Guizhou University:Natural Sciences

基　　金：贵州省档案局项目(2015D002)

摘　　要：SVD++算法由于其能够融入评分信息和隐式信息得到了广泛的应用。SVD++算法中的模型参数可以通过随机梯度下降或者最小二乘法得到最佳参数。但是SVD++中的超参数需要手动选取,而超参数会极大影响SVD++模型的准确度。本文使用两个常见的数据集,并且选取了不同的训练集-测试集切割比,使用随机梯度下降算法,对SVD++中的两个超参数学习步长和规则化参数进行研究,选用评分指标中应用最广的三个评测指标RMSE、MAE、MSE,分别得到了效果最好的超参数值。同时通过实验,本文得到了超参数对SVD++算法影响规律和一系列结论,对SVD++算法在其他数据上如何选取最好超参数值有参考意义。SVD＋＋ algorithm has been widely used for its integration into scoring information and implicit information. The model parameters（ the best parameters） of the SVD＋＋ algorithm can be obtained by the random gradient descent or the least square method. However,the super parameters in SVD＋＋ need to be selected manually and the selection values of hyper parameters will greatly affect the accuracy of the SVD＋＋ model. To solve the problem,two common data sets were used and a different cutting ratio of training set to test set was selected. By using stochastic gradient descent algorithm,two super parameter learning step and regularization parameter in SVD ＋ ＋were studied. And the best hyper-parameters were derived in RMSE,MAE,MSE which are three widely-used ones among the evaluation indicators. Results show the influence rule of the super parameter on the SVD＋＋ algorithm and a series of conclusions are obtained. This is of great reference to the selection of the best hyper parameters on the SVD＋＋ algorithm in other data sets.

关 键 词：SVD++算法 推荐算法 规则化参数 学习步长 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]