融合物品相似性与流形正则化的矩阵分解推荐模型  

Matrix Factorization Model Incorporating Item-wise Similarity and Manifold Regularization

作　　者：王嘉琦[1] 顾晓梅[1] WANG Jia-qi;GU Xiao-mei(School of Foreign Languages and Cultures,Nanjing Normal University,Nanjing 210046,China)

机构地区：[1]南京师范大学外国语学院,南京210046

出　　处：《小型微型计算机系统》2021年第2期297-302,共6页Journal of Chinese Computer Systems

基　　金：江苏省高校哲学社会科学研究一般项目(2019SJA0237)资助;2018年度南京师范大学教育教学改革项目资助.

摘　　要：矩阵分解是推荐系统中应用最为广泛的方法之一,但其对物品隐因子及其相似性学习不够充分.社会网络分析中认为相互连接的个体有一定共性,受此启发提出一个能够借助近邻关系有效学习物品隐因子及其相似性的矩阵分解推荐模型.首先基于评分矩阵对物品相似性计算进行改良,综合同一用户和相似用户的评分共现信息对物品信息建模;然后通过构建相似性优化和流形局部保持正则化项,使物品相似性作用在矩阵分解中,从而充分学习物品隐因子特征及其相似性;最后根据用户和物品隐因子矩阵计算推荐指数.在公开数据集上的实验结果表明,通过流形正则化技术将改良的物品相似性作用在矩阵分解中,可以有效提升推荐效果.This paper proposes a matrix factorization recommender model which incorporates improved item-wise similarity and manifold regularization to improve recommendation efficiency.We measure item-wise similarity by the intuition that items liked by the same users and friends are more likely to be similar.The rating matrix is factorized into user latent factors and item latent factors.We construct two regularization terms of similarity optimization and locality preserving in manifold learning,so that the neighborhood structure in which distances between similar items should be minor in latent space is related to item-wise similarity and introduced by manifold regularization techniques.We reconstruct the ratings by user and item latent factors to make recommendations.The experimental results suggest that our model which introduces item-wise similarity into matrix factorization by manifold regularization techniques outperforms some state-of-the-art models.It learns more about item latent features and measures the similarity accurately.

关 键 词：推荐系统 矩阵分解 流形学习 相似性度量 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]