引用本文
  • 刘佳玮,石川,杨成,菲利普·俞.基于异质信息网络的推荐系统研究综述[J].信息安全学报,2021,6(5):1-16    [点击复制]
  • LIU Jiawei,SHI Chuan,YANG Cheng,Philip S. Yu.Heterogeneous Information Network based Recommender Systems: a survey[J].Journal of Cyber Security,2021,6(5):1-16   [点击复制]
【打印本页】 【在线阅读全文】【下载PDF全文】 查看/发表评论下载PDF阅读器关闭

←前一篇|后一篇→

过刊浏览    高级检索

本文已被:浏览 2206次   下载 1258 本文二维码信息
码上扫一扫!
基于异质信息网络的推荐系统研究综述
刘佳玮1,2, 石川1,2, 杨成1,2, 菲利普·俞3
0
(1.北京邮电大学计算机学院 北京 中国 100876;2.智能通信软件与多媒体北京市重点实验室 北京 中国 100876;3.伊利诺伊大学芝加哥分校计算机科学系 芝加哥 美国 60607)
摘要:
推荐系统旨在为用户提供个性化匹配服务,从而有效缓解大数据时代的信息过载问题,并且改善用户体验,增加用户粘性,极大地促进了电子商务等领域的发展。然而,在实际应用场景中,由于数据稀疏和冷启动问题的存在,推荐系统往往难以得到精准的推荐结果;而复杂的模型设计也导致推荐系统的可解释性不尽如人意。因此,如何充分利用交互、属性、以及各种辅助信息提升推荐的性能和可解释性是推荐系统的核心问题。另一方面,异质信息网络作为一种全面地建模复杂系统中丰富的结构和语义信息的方法,在融合多源信息、捕捉结构语义等方面具有显著优势,已经被成功应用于相似性度量、节点聚类、链接预测、排序等各种数据挖掘任务中。近年来,采用异质信息网络统一建模推荐系统中不同类型对象的复杂交互行为、丰富的用户和商品属性以及各种各样的辅助信息,不仅有效地缓解了推荐系统的数据稀疏和冷启动问题,而且具有较好的可解释性,并因此得到了广泛关注与应用。本文旨在对基于异质信息网络的推荐系统进行全面地综述,首次系统地梳理现有工作,弥补该领域缺乏综述的空白。具体而言,本文首先介绍了异质信息网络和推荐系统的核心概念和背景知识,简要回顾了异质信息网络和推荐系统的研究现状,并且阐述了将推荐系统建模为异质信息网络的一般步骤。然后,本文根据模型原理的不同将现有方法分为三类,分别是基于相似性度量的方法、基于矩阵分解的方法和基于图表示学习的方法,并对每类方法的代表性工作进行了全面的介绍,指出了每类方法的优缺点和不同方法之间的发展脉络与内在关系。最后,本文讨论了现有方法存在的问题,并展望了该领域未来的几个潜在的研究方向。
关键词:  异质信息网络  推荐系统
DOI:10.19363/J.cnki.cn10-1380/tn.2021.09.01
投稿时间:2021-05-11修订日期:2021-08-08
基金项目:本课题得到国家自然科学基金(No.U20B2045,No.62002029,No.61772082,No.61702296)、国家重点研发计划(No.2018YFB1402600)资助。
Heterogeneous Information Network based Recommender Systems: a survey
LIU Jiawei1,2, SHI Chuan1,2, YANG Cheng1,2, Philip S. Yu3
(1.School of Computer Science, Beijing University of Posts and Telecommunications, Beijing 100876, China;2.Beijing Key Laboratory of Intelligent Communication Software and Multimedia, Beijing 100876, China;3.Department of Computer Science, University of Illinois at Chicago, Chicago 60607, USA)
Abstract:
Recommender systems aim to provide users with personalized matching services, so as to effectively alleviate the information overload problem in the era of big data, improve user experience, increase user stickiness, and greatly promote the development of e-commerce and other fields. However, in actual application scenarios, it is often difficult for recommender systems to obtain accurate recommendation results due to data sparseness and cold start problems, and the complicated model designs also lead to the unsatisfactory interpretability of the recommender systems. Therefore, how to make full use of interaction, attributes, and various auxiliary information to improve the performance and interpretability of the recommendation is the core issue of the recommender systems. For another thing, the heterogeneous information network, as a method for comprehensively modeling the rich structural and semantic information in complex systems, has significant advantages in fusing multi-source information and capturing structural semantics, and has been successfully applied to various data mining tasks such as similarity measurement, node clustering, link prediction, ranking, etc. In recent years, heterogeneous information network has been used to uniformly model the complex interactive behaviors, rich user and item attributes and various auxiliary information of different types of objects in the recommender systems, which not only effectively alleviate data sparsity and cold start problems lying in the recommender systems, but also have good interpretability and have received widespread attention and applications for the above reasons. This article aims to provide a comprehensive overview of recommender systems based on heterogeneous information networks, systematically combs the existing works for the first time and fills in the blank of the lack of review in this field. Specifically, this article first introduces the core concepts and background knowledge of heterogeneous information networks and recommender systems, briefly reviews the research status of heterogeneous information networks and recommender systems, and expounds the general steps of modeling recommender systems as heterogeneous information networks. Then, this article divides the existing methods into three categories based on the different model principles, which are methods based on similarity measurement, methods based on matrix decomposition, and methods based on graph representation learning, and gives a comprehensive introduction to the representative works of each type of methods, points out the advantages and disadvantages of each type of methods and the development context and inner relationship between different methods. Finally, this article discusses the problems of existing methods and looks forward to several potential future research directions in this field.
Key words:  heterogeneous information network  recommender system