信任网络驱动的推荐算法深度解析

需积分: 9 185 浏览量更新于2024-07-19 收藏 5.83MB PDF 举报

"《信任网络在推荐系统中的应用》是一本专门聚焦于置信推荐算法的著作，它在众多推荐系统文献中独树一帜。本书由Atlantis Computational Intelligence Systems系列编辑Daruang主持，该系列由比利时核研究中心(SCK•CEN)和根特大学（Mol & Ghent University）合作出版，旨在呈现计算智能领域的最新研究成果和发展动态。Atlantis Computational Intelligence Systems系列关注的是计算机智能的基础理论和前沿进展，包括基础研究和应用研究，以及源自计算智能研究的新技术和创新应用，涵盖范围广泛，从工业应用到生命科学的科研项目，如生物学、物理学、化学和神经科学等领域。书中探讨的信任网络在推荐系统中扮演着关键角色，它考虑了用户之间的信任度、评价和交互，通过建立用户间的信任关系来改进推荐的准确性和用户满意度。这种算法利用了社交网络分析和协作过滤技术，不仅根据用户的直接喜好，还考虑到他们的社会影响力和对其他用户推荐的信任程度。例如，一个用户可能会因为对某个朋友的信任而更倾向于采纳其推荐的产品或服务，而不是仅仅依赖于个人的历史行为数据。信任网络推荐算法的优势在于它能够解决冷启动问题，即新用户或新产品没有足够的历史数据时，如何进行推荐。通过信任关系，系统能够找到相似度较高的用户群，从而提供更为个性化的推荐。此外，信任网络模型还能有助于减少推荐的过度拟合和提高系统的稳定性，因为它在一定程度上降低了对用户行为数据的依赖。然而，设计和实施这样的信任网络并非易事，需要考虑的因素包括信任度的量化、网络结构的优化、以及如何平衡推荐的多样性与准确性。本书可能详细讨论了这些挑战以及相应的解决方案，同时提供了实用的案例研究和评估方法，以帮助读者深入理解并应用信任网络在推荐系统中的实践价值。《Trust network for Recommendation》是一本不可或缺的参考资料，对于理解和开发现代推荐系统中基于信任的策略，无论是学术研究人员还是业界从业者，都将从中受益匪浅。阅读这本书将有助于读者拓宽视野，提升推荐系统的性能，并推动计算机智能在推荐领域的发展。"

Introduction 3

a user connected to x is reached. The process of predicting the trust score along the thus

constructed path from a to x is called trust propagation. Since it often happens that a has

not one, but several trust connections that it can consult for an opinion on x , we also require

a mechanism for combining several trust scores originating from different sources. This

process is called trust aggregation. Propagation and aggregation are the two key building

blocks of trust metrics, a set of techniques which aim to estimate the trust between two

unknown users in the network. Note that the word metric has a different meaning here than

the traditional mathematical notion of metric as distance measure in a metric space.

So far, we have not mentioned context and/or goal, although this is an important factor

in computing trust estimations and taking decisions based on them: for example, your

neighbor might be very good at ﬁxing bicycles, so naturally you would trust him with your

ﬂat tire, but that does not imply that you would trust him to baby-sit your six months old

daughter. A lot of trust frameworks take into account the trust context, especially when

they are general models to be used in several kinds of applications/networks, see e.g. [1,

61]. In this book, we omit the context factor for the sake of simplicity (chapters 2-4) and

because we focus on recommendation systems for one type of items only; in other words,

we work on recommendations and trust statements that belong to the same domain/context

(chapters 6-7).

Omitting the context factor does not harm generality: while an agent in a simpliﬁed trust

model without context can choose between one or more trust statement types (e.g. trust and

distrust in a binary application such as Epinions, or 6 gradual trust levels in CouchSurﬁng),

in a trust application that takes into account context, each one of these possible statements

must be accompanied by a context statement. In this respect, we can see a trust connection

between two agents as a couple (trust statement,context statement). The operators of chap-

ters 3 and 4 can then easily be applied in context-aware trust applications as well, since

propagation and aggregation can only be performed sensibly on trust estimations within

the same context.

Obviously, context and goal are also present when deﬁning trust. For example, Gambetta

deﬁnes trust as a subjective probability that an agent will perform a particular action which

affects his own actions [31], and Jøsang et al. as the extent to which one is willing to de-

pend on somebody in a given situation [63]. As trust is used in a wide range of application

domains, plenty of trust deﬁnitions exist. Many of them focus on a different aspect of trust,

or stem from a different background (e.g. social sciences versus agent theory): Mui et al.

see trust as a subjective expectation about an agent’s future behavior based on the history of

4 Trust Networks for Recommender Systems

encounters between the agents [99], while Castelfranchi and Falcone augment Gambetta’s

deﬁnition with a competence dimension [17]. These examples illustrate that there is no

consensus on how to deﬁne trust. In this book, in which we focus on trust that is explicitly

given by the users of a recommender application domain, we adopt the general deﬁnition of

Jøsang et al. We will deﬁne trust more precisely when needed to grasp the rationale behind

trust operators and their properties.

1.2 Recommender Systems

In the previous pages, we mentioned a lot of Web 2.0 applications: Facebook, MySpace,

Twitter, Last.fm, blogs, wikis,... For the remainder of this work, however, we will focus

on one speciﬁc type of applications, namely social recommender systems. Social recom-

mender systems use information about their user’s proﬁles and/or relationships to sug-

gest items that might be of interest to them [121]. Such suggestions (recommendations)

can come in many forms: top 10 lists, promotions, ‘people who liked movie x also liked

movie y’, ‘80% of all people found the following review helpful’, etc. And it is certainly

not all about movies, cds or books only; also other ﬁelds might beneﬁt from a good rec-

ommendation system; think for example of research papers, travel packages, courses, and

so on.

Good and accurate recommender applications that guide users through the vast amounts of

online information are gaining tremendous importance, as the wealth of information makes

it increasingly difﬁcult to ﬁnd exactly what you want or need; all the more because every

person has his own preferences. Suppose that you want to go to the movies, but have no

idea what to choose: you can surely ﬁnd a lot of opinions and reviews online, but how do

you know which ones are the closest to your tastes and likes? This is where personalized

recommendation systems come into play.

From an e-commerce perspective too, the value of a good recommender system cannot be

underestimated: Cinematch, the recommender of the American online movie rental system

Netﬂix

, delivers two third of Netﬂix’s rented movies, Google News

recommendations

generate 38% click-throughs, and Amazon.com claims that 35% of their sales results from

recommendations [72]. Their importance is even more illustrated by the Netﬂix prize com-

petition, which offered a $ 1 000 000 reward for any recommendation algorithm that is 10%

See www.netflix.com and news.google.com

Introduction 5

more accurate than their own Cinematch

Most widely used recommendation systems are either content-based or collaborative ﬁlter-

ing methods. Content-based systems tend to have their recommendation scope limited to

the immediate neighborhood of a user’s past purchase or rating record. For instance, if you

have highly rated a romantic movie with Keanu Reeves, your next recommendation might

be a romantic movie or a movie featuring Keanu. The system will continue to recommend

related items only, and not explore your other interests. In this sense, recommender sys-

tems can be improved signiﬁcantly by (additionally) using collaborative ﬁltering, which

typically identiﬁes users whose tastes are similar to yours (we call them ‘neighbors’) and

recommends items that they have liked. This technique allows for more serendipitous rec-

ommendations: you might receive recommendations for movies in a genre that you are not

familiar with but that are appreciated by your neighbors, so that there is a good chance that

you will like them too.

The advanced recommendation techniques that we will discuss in this book adhere

to the collaborative ﬁltering paradigm, in the sense that a recommendation for an item

is based on ratings by other users for that item, rather than on an analysis of the item’s

content. In this sense, as with collaborative ﬁltering systems, they also belong to the class

of social recommender systems. More speciﬁcally, we will focus on one present-day set of

social recommenders, namely trust-enhanced recommender systems. The social dimension

reaches a whole new level, since trust-enhanced recommenders mine the trust network

among their users to offer their services. Such systems incorporate a trust network in which

the users are connected by scores indicating how much they trust and/or distrust each other,

and use that knowledge to generate recommendations: users can receive recommendations

for items rated highly by people in their web of trust (WOT), or even by people who are

trusted by these WOT members (through trust propagation and aggregation) etc., yielding

more, more accurate, and more personalized recommendations.

1.3 Overview

Trust metrics and recommendation technologies constitute the two pillars of trust-enhanced

recommender systems. Trust metrics are covered in detail in chapters 2-4, while chapter 5

deals with the basics of recommender systems. In chapters 6-7 we focus on the intersection

of the two ﬁelds, viz. trust-enhanced recommender systems.

See http://www.netflixprize.com

6 Trust Networks for Recommender Systems

In chapter 2 we give an overview of existing trust and distrust models and explain their

shortcomings. Current models are either not capable of properly handling inconsistency,

or cannot differentiate unknown agents from malicious agents. These shortcomings can

possibly have a large effect on the (ranking) of trust estimations, recommendations, etc.

Therefore, to meet the needs for a framework that can help agents to make better informed

(trust) opinions, we propose a new bilattice-based model that preserves valuable prove-

nance information including partial trust, partial distrust, ignorance and inconsistency.

The following two chapters focus on the mechanisms that are needed to predict trust

and distrust values in this framework. Chapter 3 covers the propagation problem. Whereas

there is a general consensus on how trust can be propagated, the picture gets much more

complicated when also distrust is involved. We describe the state of the art of trust propa-

gation, and embark upon the problem of distrust propagation, a research area that has not

received much attention so far. We discuss possible distrust propagation strategies, propose

and examine a set of propagation operators that exhibit the desired behavior, and illus-

trate them by investigating propagation patterns in real-world data sets from Epinions and

CouchSurﬁng.

Chapter 4 concentrates on aggregation techniques for trust and distrust values. This

ﬁeld, too, is still in its very infancy. To help in reaching a better understanding of the prob-

lem, we propose a set of properties that aggregation operators should fulﬁll in a (dis)trust

context. We demonstrate that the classical aggregation operators for bilattice elements are

not always suitable, and therefore propose new families of aggregation operators for trust-

enhanced applications. We examine their behavior and show their applicability on data sets

from CouchSurﬁng and Epinions.

The second part of the book deals with the application of trust metrics and their operators

in the ﬁeld of recommender systems. In chapter 5 we cover the recommender basics which

are vital for a good understanding of the subsequent chapters. We explain the collabora-

tive ﬁltering mechanism and discuss common evaluation methods and measures (related

to coverage and accuracy). We examine the problems of classical recommendation sys-

tems — transparency, sparsity, malicious users, cold start users, controversial items — and

propose a new detection measure for the latter, which is more suited for evaluation of the

corresponding shortcoming.

In chapter 6, we focus on trust- and distrust-enhanced recommendation systems, and

show how they can alleviate the problems pointed out in the previous chapter. We provide

剩余209页未读，继续阅读

darksnipers

粉丝: 0
资源: 10

信任网络驱动的推荐算法深度解析

2018-Collaborative Memory Network for Recommendation Systems.pdf

KGAT: Knowledge Graph Attention Network for Recommendation

heterogeneous graph neural network for recommendation

A matrix factorization technique with trust propagation for recommendation in social networks

”Heterogeneous Interactive Snapshot Network for Review-Enhanced Stock Profiling and Recommendation“你能给我介绍下这篇文章吗

Time Interval Aware Self-Attention for Sequential Recommendation模型解析

Dual graph convolutional network

Efficient Video Recommendation with Multi-Head Self-Attention and Hybrid Sampling

在麦吉尔大学的graduation approval query系统中存在如下信息： 1st Major Deficiencies Recommendation Yes Faculty Approval for Graduation Pending，请问这是什么意思

beat edit formac

最新资源