基于深度矩阵分解网络的矩阵填充方法  

Deep Matrix Factorization Network for Matrix Completion

作　　者：邝神芬 黄业文[3] 宋杰 李洽 KUANG Shen-fen;HUANG Ye-wen;SONG Jie;LI Qia(School of Mathematics and Statistics,Shaoguan University,Shaoguan,Guangdong 512005,China;School of Data and Computer Science,Sun Yat-sen University,Guangzhou 510006,China;School of Computer Engineering,Guangzhou College of South China University of Technology,Guangzhou 510800,China)

机构地区：[1]韶关学院数学与统计学院,广东韶关512005 [2]中山大学数据科学与计算机学院,广州510006 [3]华南理工大学广州学院计算机工程学院,广州510800

出　　处：《计算机科学》2019年第10期55-62,共8页Computer Science

基　　金：韶关学院校级项目(SY2016KJ17,SY2016KJ04);广东省教育厅青年创新人才项目(2018KQNCX233),广东省教育厅基础研究和应用基础研究重点项目(2018KZDXM065);广东省自然科学基金(2017A030307022,2018A0303100015);华南理工大学广州学院优秀骨干教师科研项目(56-CQ18YG18)资助

摘　　要：矩阵分解是矩阵填充中的流行方法,但现有的方法大多是基于浅层的线性模型,当数据矩阵变大且观测数据很少时,容易导致过拟合,性能也随之显著下降。针对这些问题,提出了一种基于深度矩阵分解网络(DMFN)的矩阵填充方法,该方法不仅能弥补传统矩阵分解的缺点,而且能处理复杂的非线性数据。首先,将输入矩阵的观测值对应的行和列向量作为输入,对其进行投影,得到其行(列)的潜在特征向量;然后,分别对行(列)的潜在特征向量构建多层感知器网络;最后,通过构建双线性池化层,将行和列的输出向量进行融合。在推荐系统数据集MovieLens及Netflix上进行测试,实验结果表明,在相同参数设置下,与主流的填充算法相比,所提方法填充预测的均方误差(RMSE)及绝对值误差(MAE)都有明显提高。Matrix factorization is a popular technique for matrix completion,but most of the existing methods are based on linear or shallow models,when the data matrix becomes large and the observation data is very small,it is prone to overfitting and the performance decreases significantly.To address this problem,this paper presented a Deep Matrix Factorization Network (DMFN) method,which can not only overcome the shortcoming of traditional matrix factorization,but also deal with complex non-linear data.First,by using rows and columns vectors corresponding to the observed values of the input matrices as input,the latent vector of its row (column) is projected,then the multi-layer perceptron network is constructed on the latent vector of row (column),at last the output vectors of rows and columns are fused by the bilinear pool layer.The proposed algorithm was tested on the standard recommender system dataset (MovieLens and Netflix).Experimental results show that the mean square error (RMSE) and mean absolute error (MAE) of the proposed method are significantly improved compared with the current popular methods under the same parameter setting.The RMSE is 0.830 and MAE is 0.652 on the MovieLens 1M dataset,which increases by about 2% and 6%,respectively.On the Netflix dataset,RMSE is 0.840 and MAE is 0.661,which increases by approximately 1% and 4%,respectively,and achieves optimal results.

关 键 词：矩阵填充 矩阵分解 深度学习 双线性池化 多层感知器 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]