CASIA OpenIR  > 毕业生  > 硕士学位论文
Alternative TitleResearch and application of Tag-based recommendation algorithm based on reinforcement learning
Thesis Advisor张文生
Degree Grantor中国科学院研究生院
Place of Conferral中国科学院自动化研究所
Degree Discipline计算机应用技术
Keyword强化学习 函数近似 推荐 标签 协同过滤 Reinforcement Learning Function Approximation Recommendation Tag Collaborative Filtering
Abstract在现实社会中,信息数量正飞速增长,在海量数据面前,人们经常会感到困惑,因此人们对个性化服务要求越来越高,这使得个性化服务发展越来越快。在个性化服务中,个性化系统对推荐算法的要求越来越高,考虑到机器学习中强化学习强劲的应用背景,本文引进强化学习思想,探索性地解决推荐技术中存在的一些问题。 在本文中,首先简要介绍了目前个性化推荐系统的研究意义以及主要研究现状;其次介绍了强化学习基本理论和经典算法,包括:TD-learning,Q-learning,资格迹等,并对函数近似在强化学习中的应用做了较为详细的介绍,讨论了各种算法之间的联系,区别及适用领域;再次介绍了当前个性化推荐中应用最成功的协同过滤推荐算法,并分析协同过滤推荐算法中存在的主要问题和对应的解决方案: 1. 数据稀疏性问题,由于历史用户的资源访问数据较少,协同过滤算法难以获得足够的分析数据,导致协同过滤推荐算法中的计算难以获得较好的准确度。本文引入目前多数推荐系统中用户对商品标注的个性化标签数据(个性化标签数据代表了用户对商品的描述信息),利用标签数据对用户的兴趣进行模拟来构建非稀疏的个性化数据。 2. 用户兴趣迁移问题,现实生活中用户兴趣容易发生迁移,然而协同过滤推荐算法将所有访问数据平等对待,所以协同过滤往往会受到用户兴趣迁移的影响。本文在构建用户兴趣模型时,考虑了用户兴趣迁移问题,对用户最新的访问数据在训练时赋予更高的权重。 在此基础上,本文设计了基于标签的强化学习推荐模型-TIRLR,将大量历史用户访问数据和基于标签模拟的用户个性化数据有效结合。并且在强化学习框架下有效的学习了用户的兴趣模型。最后实验结果表明,TIRLR相比于传统经典的基于资源的协同过滤推荐算法、基于用户的协同过滤推荐算法,均能取得更优的效果。
Other AbstractIn real society, the amount of information is growing at a rapid speed, in the face of massive data, people often feel confused, so people are demanding personalized service more urgently, and this make the development of personalized service faster. Requirements of personalization system on the recommendation algorithm are increasing. Considering strong application background of reinforcement learning in machine learning, this paper explored the introduction of ideas of reinforcement learning techniques to solve some problems in recommendation system. In this paper, we first introduce the significance and major Research in current personalized recommendation systems. Secondly we made a description on the basic theory and classical reinforcement learning algorithms, this included: TD-learning, Q-learning, eligibility traces, etc., and to do a detailed description on function approximation used in reinforcement learning, discussed links, useable areas and differences between different algorithms. Then introduced most successful application named collaborative filtering algorithm in current personalized recommendations, analyzed the main problems in collaborative filtering algorithm and solving methods. a) Due to scarce history of user access to data resources, collaborative filtering algorithms are often difficult to obtain adequate analysis data, this will cause collaborative filtering algorithm to obtain good calculation accuracy in difficulty. This paper made introduction of individual commodities tagging data in the current recommendation system, these tag data provided very useful descriptions to users, we used these tags to simulate the user's interest to build non-sparse data. b) As the migration of interest occurs to user, however, historical data will be treated equal in collaborative filtering, so the collaborative filtering tend to be impacted by interest migration. We considered the user's interest migration while modeling the user, gave greater weight to latest data while training user interest model. On the basis, we designed a tag informed reinforcement learning recommendation model -TIRLR, the large amount of historical data and tag-based user access to the user personalized data can be combined effectively. We made effective learning of the user's interest model in reinforcement learning framework. Experimental results show that, compared to classical traditional resource-based collaborative filtering algorithm and user...
Other Identifier200728017029212
Document Type学位论文
Recommended Citation
GB/T 7714
李益群. 基于标签的强化学习推荐算法研究与应用[D]. 中国科学院自动化研究所. 中国科学院研究生院,2010.
Files in This Item:
File Name/Size DocType Version Access License
CASIA_20072801702921(1060KB) 暂不开放CC BY-NC-SAApplication Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[李益群]'s Articles
Baidu academic
Similar articles in Baidu academic
[李益群]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[李益群]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.