引用本文
  • 阳德青,夏西,叶琳,薛吕欣,肖仰华.知识驱动的推荐系统:现状与展望[J].信息安全学报,2021,6(5):35-51    [点击复制]
  • YANG Deqing,XIA Xi,YE Lin,XUE Lyuxin,XIAO Yanghua.Knowledge-enhanced Recommender Systems: A Survey and Prospect[J].Journal of Cyber Security,2021,6(5):35-51   [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 1840次   下载 1218 本文二维码信息
码上扫一扫!
知识驱动的推荐系统:现状与展望
阳德青1,2, 夏西1, 叶琳1, 薛吕欣1, 肖仰华3,2
0
(1.复旦大学大数据学院 上海 中国 200433;2.复旦爱数认知智能联合研究中心 上海 中国 200433;3.复旦大学计算机科学技术学院 上海 中国 200433)
摘要:
个性化推荐系统能够根据用户的个性化偏好和需要,自动、快速、精准地为用户提供其所需的互联网资源,已成为当今大数据时代应用最广泛的信息检索系统,具有巨大的商业应用价值。近年来,随着互联网海量数据的激增,人工智能技术的快速发展与普及,以知识图谱为代表的大数据知识工程日益受到学界和业界的高度关注,也有力地推动推荐系统和个性化推荐技术也迈入到知识驱动与赋能的发展阶段。将知识图谱中蕴含的丰富知识作为有用的辅助信息引入推荐系统,不仅能够有效应对数据稀疏、语义失配等传统推荐系统难以避免的问题,还能帮助推荐系统产生多样化、可解释的推荐结果,并更好地完成跨领域推荐、序列化推荐等具有挑战性的推荐任务,从而提升各类实际推荐场景中的用户满意度。本文将现有融入知识图谱的各种推荐模型按其采用的推荐算法与面向的推荐场景不同进行分类,构建科学、合理的分类体系。其中,按照推荐方法的不同,划分出基于特征表示的和基于图结构的两大类推荐模型;按推荐场景划分,特别关注多样化推荐、可解释推荐、序列化推荐与跨领域推荐。然后,我们在各类推荐模型中分别选取代表性的研究工作进行介绍,还简要对比了各个模型的特点与优劣。此外,本文还结合当下人工智能技术和应用的发展趋势,展望了认知智能推荐系统的发展前景,具体包括融合多模态知识的推荐系统,具有常识理解能力的推荐系统,以及解说式、劝说式、抗辩式推荐系统。本文的综述内容和展望可作为推荐系统未来研究方向的有益参考。
关键词:  推荐系统  知识图谱  嵌入表示  深度学习  认知智能
DOI:10.19363/J.cnki.cn10-1380/tn.2021.09.03
投稿时间:2021-05-07修订日期:2021-08-09
基金项目:本课题得到上海市“科技创新行动计划”人工智能科技支撑专项项目(No.21511100400,No.19511120400)资助。
Knowledge-enhanced Recommender Systems: A Survey and Prospect
YANG Deqing1,2, XIA Xi1, YE Lin1, XUE Lyuxin1, XIAO Yanghua3,2
(1.School of Data Science, Fudan University, Shanghai 200433, China;2.FUDAN-AISHU Recognitive Intelligence Research Center, Shanghai 200433, China;3.School of Computer Science, Fudan University, Shanghai 200433, China)
Abstract:
Personalized recommender systems can provide users with their expected internet resources automatically, rapidly and precisely, based on users' preferences and demands. As the broadly applied information retrieval systems in the big data era, personalized recommender systems have shown great commercial values in various applications. In recent years, the knowledge engineering of big data which takes knowledge graph (KGs) as representative, has become the hot spot in the communities of both industry and academic, thanks to the sharp increasement of the big data on Web, as well as the rapid development and popularization of artificial intelligence technologies. It also promotes recommender systems to step into the phase of being driven and enabled by knowledge. With the rich knowledge in KGs used as valuable side information, the inevitable problems harming traditional recommender systems, such as data sparsity and semantic mismatch, can be addressed well. Moreover, the knowledge is helpful to generate diverse and explainable recommendation results, as well as achieve cross-domain and sequential recommendation. As a result, user experiences in various real recommendation scenarios are enhanced. In this paper, we classify the existing recommendation models incorporating KGs at first, according to their algorithm design and applied recommendation scenarios (tasks), to construct a scientific and reasonable taxonomy. Specifically, according to recommendation algorithm design, we divide the KG-based recommendation models into the class based on feature representation and the class based on graph topology. For recommendation scenarios, we specially focus on diversified recommendation, explainable recommendation, sequential recommendation and cross-domain recommendation. Next, we introduce some representative works respectively for each category, and briefly compare their characteristics including advantages and disadvantages. Furthermore, inspired by the current trend of artificial intelligence technologies and various applications, we also try to forecast the development trend of recommender systems with cognitive intelligence. To be specific, these future recommender systems include the systems built with multimodal knowledge and common sense, as well as the explainable, persuasive and polemical recommender systems. The survey and prospect in this paper can provide constructive preferences for future work in the research field of recommender systems.
Key words:  recommender system  knowledge graph  embedding  deep learning  cognitive intelligence