基于深度学习的推荐系统研究综述.pdf

计 算 机 学 报 2018 年发布 CHINESE JOURNAL OF COMPUTERS 2018Online ——————————————— 本课题得到国家重点基础研究发展计划（973） （Nos. 2014CB340401）； 国家自然科学基金重大研究计划 （Nos. 91638301）； 国家自然科学基金（Nos. 61272111， 61273216， 6160011950） 资助. 黄立威(通信作者)， 男， 1985年生， 博士， 中国计算机学会会员， 主要研究方向为机器学习、 推荐系统.E-mail: dr_huanglw@. 江碧涛，女，1967年生，博士，研究员，主要研究方向为数据挖掘. 吕守业，男，1979年生，博士，研究员，主要研究方向为数据挖掘.E-mail: lvshouye@. 刘艳博，女，1988年生，硕士，主要研究方向为图像处理、机器学习.E-mail:liuyanbonudt@. 李德毅，男，1944年生，博士，研究员，中国工程院院士，主要研究方向为人工智能. 基于深度学习的推荐系统研究综述 黄立威1), 江碧涛1), 吕守业1)，刘艳博1), 李德毅2) 1)(北京市遥感信息研究所 北京 100192) 2)(清华大学计算机科学与技术系 北京 100084) 摘 要 深度学习是机器学习领域一个重要研究方向，近年来在图像处理、自然语言理解、语音识别和广告等领域取得了突破性进展。

将深度学习融入推荐系统中， 研究如何整合海量的多源异构数据， 构建更加贴合用户偏好需求的用户模型，以提高推荐系统的性能和用户满意度， 成为基于深度学习的推荐系统的主要任务 本文对近几年基于深度学习的推荐系统研究进展进行综述，分析其与传统推荐系统的区别以及优势，并对其主要的研究方向、应用进展等进行概括、比较和分析最后，对基于深度学习的推荐系统的未来发展趋势进行分析和展望 关键词 推荐系统；深度学习；协同过滤；个性化服务；数据挖掘；多源异构数据；综述 中图法分类号 TP18 Survey on Deep Learning Based Recommender Systems HUANG Li-Wei1) JIANG Bi-Tao1) LV Shou-Ye1) LIU Yan-Bo1) LI De-Yi2) 1)(Beijing Institute of Remote Sensing, Beijing 100192) 2)(Department of Computer Science and Technology, Tsinghua University, Beijing 100084) Abstract With the ever-growing volume, complexity and dynamicity of online information, recommender systems have been an effective key solution to handle the increasing information overload problem by retrieving the most relevant information and services from a huge amount of data, and providing personalized recommendation. In recent years, deep learning technology has become an important research direction in the field of machine learning, which has been widely applied in the image processing, natural language understanding, speech recognition and online advertising. Meanwhile, recent studies also demonstrate its effectiveness in coping with information retrieval and recommendation tasks. Applying deep learning techniques into recommender systems has been gaining momentum due to its state-of-the-art performances and high-quality recommendations. In this paper, we investigate the deep learning based recommender systems, for which the main tasks are how to organize the massive multi-source heterogeneous data, build more suitable user models according to user preferences requirements, and improve the performance and user satisfaction. For specific, we first introduce the basic concepts and methods of traditional recommendation systems, including content-based recommendation method, collaborative filtering and hybrid recommendation method, and then we give an overview of the main deep learning techniques and briefly introduce their applications in the recommender systems. And secondly, we provide a comprehensive summary of current research on deep learning based recommender systems. According to the data sources used in recommender systems and the classification of the traditional recommender systems, we categorize the current research into five main directions: the application of 计算机学报2 计 算 机 学 报 2018 年 deep learning in content-based recommender systems, the application of deep learning in collaborative filtering, the application of deep learning in hybrid recommender systems, the application of deep learning in social network-based recommender systems, and the application of deep learning in context-aware recommender systems. Then, we analyze the differences and advantages of deep learning based recommender systems compared with the traditional recommender systems. First, by using deep learning, complex feature engineering can be avoided, especially when faced with unstructured data such as image and video. Second, deep learning can learn the multi-level and abstract feature representation of users and items, and is able to effectively capture the non-linear and non-trivial user-item interactions. Third, deep learning can incorporate various multi-source heterogeneous data into recommender systems, and help to mitigate the data sparseness problem considerably. Finally, this paper summarizes the future development trend of deep learning based recommender systems, e.g., the combination of deep learning and traditional recommendation methods, the application of deep learning in cross domain recommendation, the combination of the attention mechanism and deep learning based recommender systems, new deep learning recommendation architectures and the interpretability of deep learning based recommender systems. In short, deep learning has become popular in the recommender systems community both in academia and in industry. Meanwhile, this area of research is very young, there is much room for improvement in the aforementioned research directions, but we also believe that deep learning will revolutionize the recommender systems dramatically and bring more opportunities in reinventing the user experiences for better customer satisfaction in the near future. Key words recommender systems; deep learning; collaborative filtering; personalized servic。