USST_Arts_112480755用户负面评价信息对协同过滤推荐算法的影响研究

VIP免费

3.0 赵德峰 2024-11-11 4 4 2.27MB 57 页 15积分

侵权投诉

摘要

随着信息技术的进步和 web2.0 的迅猛发展，互联网在给人们带来极大便利的

同时，也把过量的信息呈递到每个人的面前，人们面临信息过载的困境。在此种

情况下，推荐系统作为一种能够根据用户兴趣，向用户推荐其可能感兴趣的产品

或服务的推荐引擎应运而生，在以电子商务为代表的互联网站中得到广泛应用，

众多学者也提出了很多优秀的推荐算法，促进了个性化推荐的发展。

本文首先介绍了当前个性化推荐算法的发展和研究现状以及在研究中经常用

到的四个数据集和常用评价指标，然后分别叙述了物理学中的物质扩散和热传导

理论在个性化推荐中的应用以及基于这两种算法的混合算法。最后，本文主要进

行了以下两方面的工作：

1. 探究推荐系统中用户负面评价信息的影响和意义。用户负面评价信息在个

性化推荐算法的研究中经常处于尴尬的地位，它们或者被同化，或者由于打分低

而被抛弃，这对于本就稀少的用户打分信息来说无疑是雪上加霜。我们的研究表

明，用户负面评价信息在 MovieLens 和Netflix 数据集中所占比例不足 20%，人们

更倾向于进行正面的评价，这更表现出负面评价信息的稀有和珍贵。此外，当从

两个数据集中去掉用户的负面评价信息时，数据集的各项属性值都发生明显的变

化，而数据集属性的改变则会对推荐算法预测作用的可靠性产生一定程度的影

响。负面评价信息同时也是用户偏恶兴趣的展示，负面评价信息的利用不仅可以

提高数据利用率，而且对做出更好的推荐也具有重要意义。

2. 提出一种基于用户负面评价信息的改进推荐算法。当前数量众多的个性化

推荐算法当中，基于用户相似度的协同过滤系统是第一代被提出并得到广泛应用

的推荐系统。本文基于用户相似性的协同过滤算法，针对个性化推荐算法研究中

用户负面评价信息被同化或者忽视的问题，以用户的负面评价信息为切入点，将

此前经常被剔除掉的用户负面评价信息应用到原协同过滤算法当中，提出了基于

用户负面评价信息的新算法，接着引入九种用户相似性的计算方法来对新算法在

两个数据集上进行效果验证，数值模拟显示新算法在准确性、多样性、意外推荐

等多方面大幅提升了原推荐算法，这也意味着负面评价信息在个性化推荐中具有

十分重要的作用。

本文提出的算法从一个全新的视角和思路对个性化推荐算法和推荐系统进行

了研究，具有一定的理论意义和实际意义。鉴于用户负面评价信息的重要作用，

我们也将在进一步的研究中寻找对其更好的定义方式和利用方式，推动个性化推

荐研究的不断进步。

关键词: 个性化推荐用户-产品二部图协同过滤算法物质扩散用

户负面评价信息

ABSTRACT

With the progress of information technology and the rapid development of Web2.0,

the Internet brings people great convenience, and at the same time it can also push

excess information to the front of people which makes people faced of the information

overload problem. In such a case, recommendation system is born as a kind of

recommendation engines which can recommend objects that users may like to target

users according to users' historical interests and has been widely used in a variety of

websites represented by e-commerce. Wide application of recommendation systems

also promotes the academic research and it has been making progress driven by

numerous scholars’ research. More and more better recommendation algorithms and

systems have been proposed and then promote the development of recommendation

system.

This paper firstly analyzes the development and research status of personalized

recommendation algorithm and briefly introduces the four datasets and commonly-used

evaluation indices in personalized recommendation research. Then two applications of

physics theories in personalized recommendation are introduced including mass

diffusion and heat conduction. Finally, based on the user-object bipartite network, this

paper has mainly done the following two aspects:

1. To explore the significance and influence of users’ negative opinions in

recommendation system. Users’ negative opinions are often in an awkward position in

the research of recommendation algorithm. They are assimilated, or abandoned due to

low score, and this can be definitely icing on the cake for sparse datasets. Our research

indicates that the users’ negative opinions account for less than 20% in MovieLens and

Netflix datasets and people are more likely to make a positive rating which shows

rareness and preciousness of users’ negative opinions. In addition, when removing the

users’ negative opinions from two datasets, there is an obvious change of the attribute

values in both datasets which will have an effect on the predictive reliability of

recommendation algorithm. The negative information also shows users’ interest of

dislike, and the proper use of it can not only improve the utilization rate of data, but also

have a great significance in making better recommendation.

2. To present an improved recommendation algorithm based on users'

negative opinions. User-based collaborative filtering system is the first generation of

recommendation system which is widely used. This paper firstly introduces the concept

and importance of users’ negative information in detail through the deep analysis of

practice and data, and then proposes the recommendation algorithm based on users'

negative opinions according to the user-based collaborative filtering algorithm and the

application of users' negative information. Nine user-similarity methods are brought in

to evaluate the new algorithm in two datasets and the numerical simulation results show

that the new algorithm has a great improvement in many aspects such as accuracy,

diversity, surprising and so on compared with the primary algorithm which shows the

important role of users’ negative opinions in the personalized recommendation.

The algorithm proposed by this paper studies the personalized recommendation

algorithms and systems from a new visual angle and thought and has a certain

significance of both theory and practice. In consideration of the important role of users’

negative information, we will seek for the better way of definition and using method in

the further study and make a contribution to the development and progress of

personalized recommendation research.

Key words: Personalized recommendation, User-object bipartite

network, Collaborative filtering algorithm, Mass diffusion,

Users’ negative opinions

中文摘要

ABSTRACT

第一章绪论 ....................................................... 1

1.1 研究背景 ..................................................... 1

1.2 研究意义 ..................................................... 3

1.2.1 理论意义 ................................................ 3

1.2.2 实际意义 ................................................ 4

1.3 个性化推荐系统的发展现状和问题 ............................... 5

1.3.1 个性化推荐发展现状 ...................................... 5

1.3.2 个性化推荐发展中面临的问题 .............................. 6

1.4 本文的主要工作 ............................................... 9

1.4.1 本文的贡献 .............................................. 9

1.4.2 本文的结构 .............................................. 9

第二章复杂网络相关理论简介 ....................................... 11

2.1 复杂网络理论简介 ............................................ 11

2.2 用户-产品二部图 ............................................. 11

2.2.1 二部图简介 ............................................. 11

2.2.2 二部图相关统计属性简介 ................................. 13

第三章个性化推荐算法简介 ......................................... 16

3.1 个性化推荐算法 .............................................. 16

3.1.1 协同过滤算法 ........................................... 16

3.1.2 基于内容的推荐算法 ..................................... 18

3.1.3 基于网络结构的推荐算法 ................................. 20

3.1.4 混合推荐算法 ........................................... 21

3.2 常用数据集介绍 .............................................. 22

3.2.1 MovieLens .............................................. 22

3.2.2 Netflix ................................................ 23

3.2.3 Amazon ................................................. 24

3.2.4 Delicious .............................................. 25

3.3 算法评价指标 ................................................ 26

3.3.1 准确度 ................................................. 27

3.3.2 多样性 ................................................. 27

3.3.3 意外推荐的能力 ......................................... 27

3.3.4 去冗余的能力 ........................................... 28

第四章物理方法在个性化推荐中的应用 ............................... 29

4.1 基于热传导的个性化推荐算法 .................................. 29

4.2 基于物质扩散的个性化推荐算法 ................................ 30

4.3 混合算法 .................................................... 31

4.4 本章小结 .................................................... 32

第五章用户负面评价信息对协同过滤算法的影响研究 ................... 33

5.1 用户负面评价信息及其影响研究 ................................ 33

5.2 基于用户负面评价信息的协同过滤算法研究 ...................... 35

5.2.1 用户产品二部图和 HC 模型 ................................ 35

5.2.2 基于用户负面评价信息的协同过滤算法 ..................... 36

5.3 基于用户的相似性度量方法 .................................... 37

5.3.1 基于节点的相似性 ....................................... 37

5.3.2 基于物质扩散和热传导的用户相似性计算 ................... 38

5.4 数值结果分析 ................................................ 39

5.5 总结 ........................................................ 43

第六章结论与展望 ................................................. 45

6.1 结论 ........................................................ 45

6.2 展望 ........................................................ 45

参考文献 ........................................................... 47

在读期间公开发表论文和承担科研项目及取得的成果 ..................... 52

致谢 ............................................................ 53

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

15 积分 4人已下载

立即下载 VIP免费下载

摘要：

摘要随着信息技术的进步和web2.0的迅猛发展，互联网在给人们带来极大便利的同时，也把过量的信息呈递到每个人的面前，人们面临信息过载的困境。在此种情况下，推荐系统作为一种能够根据用户兴趣，向用户推荐其可能感兴趣的产品或服务的推荐引擎应运而生，在以电子商务为代表的互联网站中得到广泛应用，众多学者也提出了很多优秀的推荐算法，促进了个性化推荐的发展。本文首先介绍了当前个性化推荐算法的发展和研究现状以及在研究中经常用到的四个数据集和常用评价指标，然后分别叙述了物理学中的物质扩散和热传导理论在个性化推荐中的应用以及基于这两种算法的混合算法。最后，本文主要进行了以下两方面的工作：1.探究推荐系统中用户负面...

展开>> 收起<<

USST_Arts_112480755用户负面评价信息对协同过滤推荐算法的影响研究.pdf

共57页,预览6页

还剩页未读，继续阅读

USST_Arts_112480755用户负面评价信息对协同过滤推荐算法的影响研究

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

推荐作者

热门标签

举报选择: