基于二部图网络结构的个性化推荐系统研究
VIP免费
摘 要
近年来,伴随着互联网技术的迅猛发展,越来越多的用户面临着信息过载:
用户面临海量的信息以至于用户不能从中选择自己所需要的信息。人们在互联网
中获取所需信息的主要渠道仍是通过搜索引擎。推荐系统的出现,使得用户搜索
个性化的信息成为可能。推荐系统也因此受到了各个学科的普遍关注,利用复杂
网络的方法来研究推荐系统的推荐算法就是其中的一个研究方向。
利用复杂网络的方法来研究推荐系统的推荐算法,其根本思想是试图通过推
荐系统中用户选择项目的关系,构建一个二部图网络,通过二部图网络资源分配
过程,结合协同过滤算法来产生新的推荐算法。通过二部图网络资源分配过程,
可以得到一个资源分配矩阵。该矩阵代表了某种资源通过二部图结构来进行重新
分配。它可以从两个方面来考虑,其中都是假定用户选择过的项目有向用户推荐
其他项目的能力。一种算法是从项目的角度来考虑,首先是计算项目之间的相似
度,以项目的相似度在所有项目相似度中的权重作为该项目的初始推荐资源,在
向某一个用户推荐时,用户选择过的项目具有初始推荐资源,没有选择过的项目
没有初始推荐资源,然后对用户未选择过的项目进行预测打分,按照用户未选择
过项目的打分高低向用户推荐项目。第二种算法是从用户的角度来考虑,首先是
计算用户之间的相似度,以某用户相似度在所有用户相似度中的权重作为该用户
的初始推荐资源,通过资源分配矩阵重新分配,然后对用户未曾选择的项目进行
预测打分,然后向用户推荐打分最高的项目。
在验证算法的有效性以及可扩展性时,本文采用真实数据来验证。在衡量个
性化推荐系统的推荐算法时,采用的衡量指标主要有排序值,平均度以及海明距
离,相比于全局排序算法与经典的协同过滤算法,测试结果表明算法效果具有了
明显的改进。同时,本文进一步证明了基于项目的推荐算法产生的推荐效果要好
于基于用户的推荐算法。同时,针对两种算法表现的不同,从度分布以及项目
(用户)相似度同度的关系两个方面来考察产生不同结果的原因。文章最后提出
了本文的不足,以及将来进一步的研究方向。
关键词: 个性化推荐系统 协同过滤 复杂网络 二部图网络投影
ABSTRACT
With the rapid development of the Internet technology in the last decade, we
confront with information overload: there are too much data on the Internet that users
can’t access to the information they interested in. Search engineer is the main tools that
users get the information they favorite from the Internet nowadays. The emergence of
the Recommends System makes users access personal information possible. Therefore,
many disciplines began to focus on Recommendation System, and take advantage of
complex network method to solve the problem of Recommend System algorithm is one
of the research area.
The main idea of taking use of the method of complex network to solve problem of
Recommendation System algorithm is to find out the relations between the user and
item in one Recommends System and then map it into a bipartite network. And then the
resource distribution process on bipartite network is taken into account and combine
with collaborative filtering method to obtain one new recommendations method. One
resource distribution matrix representing resource distributed from one node (each item
or user maybe present by a node) to another is obtained after the process of resource
distribution on bipartite network, which is also one kind of projection from bipartite
network to unique network. It is two aspects to deal with this matrix in this paper. One
method from the item’s perspective is that the similarity between items is calculated and
the weight of the item similarity one item with any other items in system can view as
the initial recommendation resource. When the system recommend items to user, only
the item that the user already selected have the initial recommendation resource and
score the item the user unselected. Then the system recommend item to user according
to the score of the item after predication from high to low. Another method from user’s
point is that the similarity of the user is gained and the weight of the similarity one user
with any other users in this recommendation system similarly views as the initial
recommendation resource. By taking resource distribution matrix into account, the item
score that the user unselected is predicted and then the system recommend the user with
the top score one.
When verify the validation and extension of our algorithm, the real dataset is
applied to five methods motioned in this paper. There are three measurements adopted
in this paper to measure the achievement of our personal recommendation algorithm,
which is ranking value, average degree and hamming distance respectively. The
experiment result demonstrates that out method has well performance when compared
with Global Ranking method and Classical Collaborative Filtering method based on
item and user respectively. It also concluded that our method based on item is better
than our method based on user in our test dataset. Meanwhile, the degree distribution
and the relation between degree and item (user) similarity is calculated when we try to
find out the different performance of our two method. Finally, it is the shortcomings of
our method and our future work that presents.
Key words:recommendation system, collaborative filtering, complex
network, bipartite network projection
目 录
中文摘要
ABSTRACT
第一章 绪 论 ......................................................... 1
§1.1 研究背景及其意义 ........................................... 1
§1.2 研究框架 ................................................... 4
第二章 复杂网络理论简介 .............................................. 6
§2.1 单模式网络统计属性研究 ..................................... 6
§2.1.1 度及其相关属性 ........................................ 7
§2.1.2 平均最短路径 .......................................... 7
§2.1.3 聚类系数 .............................................. 8
§2.2 二部图网络投影及属性研究 ................................... 8
§2.2.1 无加权二部图投影 ...................................... 9
§2.2.2 加权方式的二部图投影 .................................. 9
§2.2.3 二部图网络属性研究 ................................... 11
§2.3 个性化推荐系统 ............................................. 14
§2.3.1 协同过滤系统 ......................................... 14
§2.3.2 基于内容的推荐系统 ................................... 17
§2.3.2 基于网络结构算法的推荐系统 ........................... 19
§2.3.3 混合推荐算法的推荐系统与推荐系统的衡量指标 ........... 20
第三章 基于项目的个性化推荐算法 ..................................... 22
§3.1 二部图网络统计属性 ........................................ 22
§3.2 推荐算法设计 .............................................. 23
§3.3 改进推荐算法 .............................................. 25
§ 3.4 仿真实验说明与算法衡量指标 ............................... 25
§ 3.5 算法结果及说明 ........................................... 26
第四章 基于用户的个性化推荐算法 ..................................... 30
§4.1 推荐算法设计 .............................................. 30
§4.2 推荐改进算法 .............................................. 31
§4.3 算法结果及说明 ............................................ 31
§4.4 小结 ...................................................... 34
第五章 论文总结和展望 ............................................... 35
摘要:
展开>>
收起<<
摘要近年来,伴随着互联网技术的迅猛发展,越来越多的用户面临着信息过载:用户面临海量的信息以至于用户不能从中选择自己所需要的信息。人们在互联网中获取所需信息的主要渠道仍是通过搜索引擎。推荐系统的出现,使得用户搜索个性化的信息成为可能。推荐系统也因此受到了各个学科的普遍关注,利用复杂网络的方法来研究推荐系统的推荐算法就是其中的一个研究方向。利用复杂网络的方法来研究推荐系统的推荐算法,其根本思想是试图通过推荐系统中用户选择项目的关系,构建一个二部图网络,通过二部图网络资源分配过程,结合协同过滤算法来产生新的推荐算法。通过二部图网络资源分配过程,可以得到一个资源分配矩阵。该矩阵代表了某种资源通过二部图...
相关推荐
作者:陈辉
分类:高等教育资料
价格:15积分
属性:49 页
大小:954.78KB
格式:PDF
时间:2024-11-19