Abstract Trust-aware Decentralized Recommender Systems

合集下载

《2024年基于多尺度和注意力机制融合的语义分割模型研究》范文

《2024年基于多尺度和注意力机制融合的语义分割模型研究》范文

《基于多尺度和注意力机制融合的语义分割模型研究》篇一一、引言随着深度学习技术的飞速发展,语义分割已成为计算机视觉领域的研究热点。

语义分割旨在将图像中的每个像素分类为预定义的语义类别,从而为自动驾驶、医疗影像分析、卫星图像解析等众多领域提供了强有力的技术支持。

近年来,多尺度和注意力机制在语义分割模型中得到了广泛应用,它们能够有效地捕获不同尺度的上下文信息,并关注重要的区域以提升分割精度。

本文将研究基于多尺度和注意力机制融合的语义分割模型,以提高模型的性能和泛化能力。

二、相关工作在语义分割领域,多尺度特征融合和注意力机制是两个重要的研究方向。

多尺度特征融合能够捕获不同尺度的上下文信息,提高模型对不同大小目标的分割精度。

而注意力机制则能关注重要的区域,抑制无关区域,从而提高模型的关注力和准确性。

近年来,许多研究工作已经将这两者结合在一起,取得了良好的效果。

三、方法本文提出了一种基于多尺度和注意力机制融合的语义分割模型。

该模型主要包括以下几个部分:1. 多尺度特征提取:采用不同尺度的卷积核和池化操作,提取多尺度的上下文信息。

这些不同尺度的特征图将作为后续模块的输入。

2. 注意力机制模块:采用自注意力机制和交叉注意力机制,对每个尺度的特征图进行加权,以关注重要的区域并抑制无关区域。

3. 特征融合与上采样:将加权后的多尺度特征图进行融合,并采用上采样操作使特征图恢复到原始图像的大小。

4. 损失函数设计:采用交叉熵损失和Dice损失相结合的损失函数,以平衡正负样本的比例并提高模型的鲁棒性。

四、实验为了验证本文提出的模型的有效性,我们在多个公开数据集上进行了实验。

实验结果表明,本文提出的模型在语义分割任务中取得了良好的效果。

具体来说,我们在Cityscapes、ADE20K 等数据集上进行了实验,并与其他先进的语义分割模型进行了比较。

实验结果显示,本文提出的模型在分割精度、速度和泛化能力等方面均有所提升。

五、结果与分析1. 性能提升:通过多尺度和注意力机制的融合,本文提出的模型在语义分割任务中取得了较好的性能提升。

深度学习的目标跟踪算法综述

深度学习的目标跟踪算法综述

深度学习的目标跟踪算法综述引言:随着深度学习技术的快速发展,目标跟踪领域也得到了巨大的发展。

目标跟踪是指在视频序列中,对感兴趣的目标进行连续的定位和跟踪,其在计算机视觉、自动驾驶、视频监控等领域有着广泛的应用前景。

本文将综述几种常见的深度学习目标跟踪算法,以便读者对这一领域有更全面的了解。

一、基于卷积神经网络的目标跟踪算法卷积神经网络(Convolutional Neural Network,CNN)是深度学习中最常用的网络结构之一。

它通过卷积层、池化层和全连接层等结构,能够自动提取图像特征。

在目标跟踪中,常用的基于CNN的算法有Siamese网络、Correlation Filter网络和DeepSORT等。

1. Siamese网络Siamese网络是一种基于孪生网络结构的目标跟踪算法,它通过输入一对图像样本来学习两个样本之间的相似度。

该网络通过训练得到的特征向量,可以用于计算待跟踪目标与骨干网络中的目标特征之间的距离,从而确定目标的位置。

2. Correlation Filter网络Correlation Filter网络是一种基于卷积神经网络的目标跟踪算法,它通过训练得到的滤波器,可以将目标与背景进行区分。

该算法通过计算滤波响应图,来确定目标的位置和尺度。

3. DeepSORTDeepSORT是一种结合深度学习和传统目标跟踪算法的方法,它通过使用CNN进行特征提取,并结合卡尔曼滤波器对目标进行预测和更新。

DeepSORT在准确性和实时性上都有较好的表现,在实际应用中有着广泛的使用。

二、基于循环神经网络的目标跟踪算法循环神经网络(Recurrent Neural Network,RNN)是一种能够处理序列数据的神经网络模型。

在目标跟踪中,RNN可以考虑到目标在时间上的依赖关系,从而提高跟踪的准确性。

常见的基于RNN的目标跟踪算法有LSTM和GRU等。

1. LSTMLSTM是一种常用的循环神经网络结构,它能够有效地处理长期依赖问题。

《2024年基于PCA的人脸识别研究》范文

《2024年基于PCA的人脸识别研究》范文

《基于PCA的人脸识别研究》篇一一、引言人脸识别技术已成为现代社会中不可或缺的一部分,其广泛应用于安全监控、身份认证、人机交互等领域。

然而,由于人脸的复杂性以及各种因素的影响,如光照、表情、姿态等,使得人脸识别成为一个具有挑战性的问题。

为了解决这些问题,研究者们提出了一种基于主成分分析(PCA)的人脸识别方法。

本文旨在探讨基于PCA的人脸识别技术的研究,包括其原理、方法、实验结果及未来发展方向。

二、PCA原理及方法PCA(Principal Component Analysis)是一种常用的统计分析方法,其主要思想是将原始特征空间中的高维数据投影到低维空间中,从而减少数据的冗余性和复杂性。

在人脸识别中,PCA通过将人脸图像的高维特征向量投影到低维空间中,以实现降维和特征提取。

具体而言,PCA方法包括以下步骤:1. 数据预处理:对原始人脸图像进行灰度化、归一化等预处理操作,以便进行后续的降维和特征提取。

2. 构建协方差矩阵:根据预处理后的人脸图像数据,构建协方差矩阵。

3. 计算特征值和特征向量:对协方差矩阵进行特征值分解,得到其特征值和特征向量。

4. 选取主成分:根据特征值的大小选取前k个主成分,构成新的低维空间。

5. 投影与降维:将原始数据投影到新的低维空间中,得到降维后的数据。

三、基于PCA的人脸识别方法基于PCA的人脸识别方法主要包括以下步骤:1. 人脸检测与预处理:通过人脸检测算法从图像中提取出人脸区域,并进行预处理操作,如灰度化、归一化等。

2. 特征提取:利用PCA方法对预处理后的人脸图像进行降维和特征提取。

3. 训练与建模:将提取的特征向量输入到分类器中进行训练和建模,如支持向量机(SVM)、神经网络等。

4. 测试与识别:将待识别的人脸图像进行同样的预处理和特征提取操作后,与训练集中的数据进行比较和匹配,从而实现人脸识别。

四、实验结果与分析本文采用ORL人脸数据库进行实验,对比了基于PCA的人脸识别方法与其他方法的性能。

《2024年深度学习相关研究综述》范文

《2024年深度学习相关研究综述》范文

《深度学习相关研究综述》篇一一、引言深度学习作为机器学习的一个分支,近年来在人工智能领域中获得了显著的突破与成功。

随着数据量的不断增加以及计算能力的提高,深度学习已经逐渐成为了众多领域研究的热点。

本文将对深度学习的基本原理、研究进展以及当前主要研究方向进行综述,旨在为读者提供一个清晰、全面的认识。

二、深度学习的基本原理深度学习是指一类基于神经网络的机器学习方法,通过构建深度神经网络,实现复杂的非线性映射,使机器能够在图像识别、语音识别、自然语言处理等任务中取得卓越的表现。

深度学习的基本原理包括神经网络的构建、前向传播和反向传播等过程。

三、深度学习的研究进展自深度学习概念提出以来,其在计算机视觉、自然语言处理、语音识别等领域取得了显著的成果。

特别是随着深度神经网络的不断发展,其在各类大型比赛中的表现越来越出色。

如:在ImageNet大规模图像识别挑战赛中,基于深度学习的算法取得了历史性的突破;在语音识别领域,深度学习技术已经可以实现在不同噪音环境下的高质量语音识别;在自然语言处理领域,基于深度学习的算法实现了自然语言生成和翻译等方面的技术革新。

四、深度学习的研究方向目前,深度学习领域的研究主要集中在以下几个方面:1. 卷积神经网络:针对图像和视频处理领域,卷积神经网络已经成为了一种有效的深度学习方法。

研究者们通过不断改进网络结构、优化参数等手段,提高了其在各类任务中的性能。

2. 循环神经网络:针对自然语言处理等领域,循环神经网络的应用逐渐得到关注。

通过利用序列数据之间的依赖关系,循环神经网络在文本生成、语音识别等方面取得了显著的成果。

3. 生成式对抗网络:生成式对抗网络是一种无监督学习方法,通过生成器和判别器之间的竞争与协作,实现数据的高质量生成和增强。

在图像生成、视频生成等领域具有广泛的应用前景。

4. 迁移学习与小样本学习:随着深度学习应用场景的扩大,如何在有限的数据下进行有效的学习和预测成为了一个重要的研究方向。

大语言模型增强因果推断

大语言模型增强因果推断

大语言模型(LLM)是一种强大的自然语言处理技术,它可以理解和生成自然语言文本,并具有广泛的应用场景。

然而,虽然LLM能够生成流畅、自然的文本,但在因果推断方面,它仍存在一些限制。

通过增强LLM的因果推断能力,我们可以更好地理解和解释人工智能系统的行为,从而提高其可信度和可靠性。

首先,我们可以通过将LLM与额外的上下文信息结合,来增强其因果推断能力。

上下文信息包括时间、地点、背景、情感等各个方面,它们可以为LLM提供更全面的信息,使其能够更好地理解事件之间的因果关系。

通过这种方式,LLM可以更好地预测未来的结果,并解释其预测的依据。

其次,我们可以通过引入可解释性建模技术,来增强LLM的因果推断能力。

这些技术包括决策树、规则归纳、贝叶斯网络等,它们可以帮助我们更好地理解LLM的决策过程,从而更准确地预测其结果。

此外,这些技术还可以帮助我们识别因果关系的路径,从而更深入地了解因果关系。

最后,我们可以通过将LLM与其他领域的知识结合,来增强其因果推断能力。

例如,我们可以将经济学、心理学、社会学等领域的知识融入LLM中,以帮助其更好地理解和解释因果关系。

通过这种方式,LLM可以更全面地考虑各种因素,从而更准确地预测和解释因果关系。

在应用方面,增强因果推断能力的LLM可以为许多领域提供更准确、更可靠的决策支持。

例如,在医疗领域,它可以辅助医生制定更有效的治疗方案;在金融领域,它可以辅助投资者做出更明智的投资决策;在政策制定领域,它可以为政策制定者提供更全面、更准确的政策建议。

总之,通过增强大语言模型(LLM)的因果推断能力,我们可以更好地理解和解释人工智能系统的行为,从而提高其可信度和可靠性。

这将有助于推动人工智能技术的广泛应用和发展,为社会带来更多的便利和价值。

同时,我们也需要关注和解决相关伦理和社会问题,以确保人工智能技术的发展符合人类的价值观和利益。

《2024年基于多尺度和注意力机制融合的语义分割模型研究》范文

《2024年基于多尺度和注意力机制融合的语义分割模型研究》范文

《基于多尺度和注意力机制融合的语义分割模型研究》篇一一、引言随着深度学习技术的不断发展,语义分割作为计算机视觉领域的一个重要任务,逐渐成为研究的热点。

语义分割旨在将图像中的每个像素划分为不同的语义类别,为图像理解提供了更加细致的信息。

然而,由于实际场景中存在多尺度目标和复杂背景的干扰,语义分割任务仍面临诸多挑战。

为了解决这些问题,本文提出了一种基于多尺度和注意力机制融合的语义分割模型。

二、相关工作语义分割作为计算机视觉的一个关键任务,在近几年的研究中得到了广泛的关注。

目前主流的语义分割模型主要采用深度卷积神经网络(CNN)来实现。

这些模型通过捕获上下文信息、提高特征表达能力等手段提高分割精度。

然而,在处理多尺度目标和复杂背景时,这些模型仍存在局限性。

为了解决这些问题,本文提出了一种融合多尺度和注意力机制的语义分割模型。

三、模型与方法本文提出的模型主要由两个部分组成:多尺度特征提取和注意力机制融合。

(一)多尺度特征提取多尺度特征提取是提高语义分割性能的关键技术之一。

在本模型中,我们采用了不同尺度的卷积核和池化操作来提取图像的多尺度特征。

具体而言,我们设计了一个包含多种尺度卷积核的卷积层,以捕获不同尺度的目标信息。

此外,我们还采用了池化操作来获取更大尺度的上下文信息。

这些多尺度特征将被用于后续的注意力机制融合。

(二)注意力机制融合注意力机制是一种有效的提高模型性能的技术,可以使得模型更加关注重要的区域。

在本模型中,我们采用了自注意力机制和交叉注意力机制来提高模型的表达能力。

自注意力机制主要用于捕获每个像素的上下文信息,而交叉注意力机制则用于融合不同尺度特征之间的信息。

具体而言,我们通过在卷积层之间引入自注意力和交叉注意力模块,使得模型能够更好地关注重要区域和提取多尺度特征。

四、实验与结果为了验证本文提出的模型的性能,我们在公开的语义分割数据集上进行了一系列实验。

实验结果表明,本文提出的模型在处理多尺度目标和复杂背景时具有更好的性能。

基于本体和上下文感知的主动式计算机犯罪取证模型设计研究

基于本体和上下文感知的主动式计算机犯罪取证模型设计研究

模型中所涉及到的实体 、策略、权限、规则都由本体语 言描述 ,
实 现 了语 义层 面的 概 念 的 规 范统 一 。
13蜜 罐 .
美 国的 L n eS i nr H ny o 定义 如下 : n y a c pt e 对 o e p t z Ho e—
上下文信息在取证 中的作用 :随着上下文信息的变化 ,用户
前 的信息状况 ,最大限度的还原犯罪 过程 。
我们使 用本 体的 目的是 用它来获 取、描 述 以及 表示相 关 领域 的知识 ,从 而提供 对该领域 知识 的共 同理解 ,对该 领域 的词汇 或术语 提供 一个共 同认可 的解 释,并从 对这些词 汇或 术语 以及词 汇或 术语问 的相 互关 系进 行明确定 义。通常情况 下,本体 具 有静 态性 和动态性 这两个 特性 ,静态 性不涉及动 态 的行 为,它反 映的是概 念模 型 ; 而动态 性是 指它 的内容 和
12本体 的概念 .
本 体是来 源于哲学领 域的一个 概念 ,近几 年来 ,计 算机 领 域的研 究者们 也开始使用 本体 的概 念 ,并将它用于模拟企
业 结 构 ,并 提 出了虚拟 企 业 这一 研 究方 向。在计 算 机领域 , 本体 的定 义经历 了一个漫长 的过程 。Grb r 19 年 给出了 u e在 9 3
堪设 想。 因此我 们把取证 系统安装 在分布式蜜 网中,保证 取
证 的安 全 。
2基于本体和 上下文感知 的动态 蜜罐取证模型
21动态取证模型 .
结合上下文和本体 ,提出了基 于本体和上下文感 知的动态
本 体 的一个 定 义,即 “ 本体是 概念 模 型的明确 的规 范说 明” 。
后来 ,B rt os 又在这个定义 的基础上作了修改,将本体定义为:

基于贝叶斯正则化神经网络的企业资信评估

基于贝叶斯正则化神经网络的企业资信评估

a d c mp t .T e e p rme tr s l h w t a h a e i n—r g lrz t n n u a ew r s s b e a d fs ,h s n o u e h x e i n e u t s o h tt e B y sa s e u aiai e r ln t o k i t l n a t a o a g o r d cin a c r c ,a d c n b p l d t h op rt rd tr t g o d p e it c u a y n a e a p i o t e c r oa e c e i ai . o e n
摘要: 在市场经济系统研究中 , 资信评估作为市场经济 中的监督力量 , 是投 资者的重要参考依据。科学准确的资信评估 可以 辅助决策 , 降低投资者风险。针对当前企业资信评估方法的不足 , 为了准确评估资信效果 , 提出将基于贝叶斯正则化 的前 向 多层神经 网络用于企业资信评估 , 通过新 的误差函数可以减少网络 的有效权值和 阈值 , 并使网络训练输出更加平滑 , 而增 从
ABS TRACT: h n t r g fr e n ma k te o o ,c e i r t g i a mp r n ee e c o n e tr .Ac Ast e mo i i oc s i r e c n my r d t ai s n i ot trf r n e f riv so s on n a — c r t n ce t cc e i r t g c n h l e iin—ma i ga d r d c e r k o v s r .T v i h n b l y o u a e a d s in i r d t ai a ep d c s i f n o k n n e u e t i f n e t s o a od t e ia i t f h s i o i c n e t n lmeh d fc r oa e ce i rt g h a e r p s st p l h e d fr r r f iln u a ewo k o v n i a t o so o p rt rd t ai ,t e p p rp o o e o a py te f e o wad a t i a e r n t r o n ic l w t a e in—r g lr ain t i i g ag r h t h r be o r d t a ig i B ys h a e u a i to r nn lo t m o t e p o lm f e i r t .T e a o t n o e e r rf n t n z a i c n h d p i f n w ro u ci o a o c n r d c h ewo k efc ie w ih sa d t r s od n k h u p t f ew r r ii g s o h r Ot a e a e u et e n t r f t eg t n h e h l sa d ma e te o tu t o k tan n mo t e ,S h t h e v on t
  1. 1、下载文档前请自行甄别文档内容的完整性,平台不提供额外的编辑、内容补充、找答案等附加服务。
  2. 2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
  3. 3、如文档侵犯您的权益,请联系客服反馈,我们会尽快为您处理(人工客服工作时间:9:00-18:30)。

Trust-aware Decentralized Recommender Systems:PhD research proposalPaolo MassaDepartment of Information and Communication Technology-University of Trento Via Sommarive14-I-38050Povo(TN)-ItalyE-mail:massa@itc.it29th May2003AbstractThis PhD thesis addresses the following problem:ex-ploiting of trust information in order to enhance the accuracy and the user acceptance of current Recom-mender Systems(RS).RSs suggest to users items they will probably like.Up to now,current RSs mainly gener-ate recommendations based on users’opinions on items. Nowadays,with the growth of online communities,e-marketplaces,weblogs and peer-to-peer networks,a new kind of information is available:rating expressed by an user on another user(trust).We analyze current RS weaknesses and show how use of trust can overcome them.We proposed a solution about exploiting of trust into RSs and underline what experiments we will run in order to test our solution.1Introduction“Although an application designer’sfirst instinct is to reduce a noble human being to a mere account number for the computer’s convenience,at the root of that account number is always a human identity”[24].We are in the Information society.The quantity of new information available every day(news,movies,sci-entific papers,songs,websites,...)goes over our lim-ited processing capabilities.For this reason,we need something able to suggest us only the worthwhile infor-mation.Recommender Systems(RS)[34,39]have this aim.In particular,RSs based on Collaborative Filter-ing(CF)[16,7]try to automate the“word of mouth”process.The intuition is the following:when we have to decide about going to see a new movie for example, we often ask to some friends with similar movies tastes and then we act based on their recommendations.CF tries to automates this process to a world scale:there is no more need the users asks to known people but it is the system(that knows the judgement of everyone) thatfinds users similar to her and recommends to her the items they like.However,RSs based on CF have some weaknesses: cold start problem[40]is related to the situation when an user enters the system and has expressed no ratings so that the RS cannotfind the like-minded users;any-way in general the quantity of ratings an user gives is very low compared to the quantity of available items(for example,Eachmovie[29]dataset is97.4%sparse),this results in a great sparseness of the data and low confi-dence of the automated strategies.Another concern is related to the existence of users that want to maliciously influence the system:if they know how the recommen-dations are built they can express fake ratings in order to achieve a certain goal(such as get their book recom-mended to everyone).There are some recent study of possible attacks in distributed systems[26,22]but we are not aware of research lines taking into account this as a problem for RSs.Moreover,with current RSs it is very hard(or impossible)for the user to control the rec-ommendation process so that if the RS starts giving bad quality recommendations,usually the user just stops in using it[46,17].Nowadays,online communities(,),e-marketplaces(,,,...),peer-to-peer networks(edonkey,gnutella) and weblogs use and make available trust information (judgement expressed by users on users).They use it with slightly different specific goals but what they share is the fact this introduce a sort of social control over the all system.We claim taking into account explicit trust informa-tion provided by users into Recommender Systems can overcome their inherent,previously cited weaknesses.We also claim this environment reclaims a decentral-ized approach:the only possible way to be in control of the recommendation process is being able to do it by yourself;in this sense,it does not make sense having one centralized server as the only entity able to access the information expressed by users and to generate rec-ommendations.In order to have a real control over the recommendation process,every single entity(we call it peer referring to the peer-to-peer architecture)must self-publish the data(ratings and trusts)so that it is possi-ble for every peer to fetch all the information and self-computes recommendation on behalf of its user.Self-publishing is made very easy by recent tools such as weblogs.We think decentralized data publishing is im-portant for research too,in fact we claim research in RS was slow and not successful as possible also because of 1lack of available datasets and real testbeds for testing innovative solutions.We propose a strategy to exploit trust information in order to enhance current RSs.Our proposed solution is a multi step one where different algorithms,caring for different types of input data,can be combined with a special emphasis to the confidence they have in the correctness of predicted information.We also present some experiments we will run in or-der to test our hypothesis.Some experiments will be on sinthesized data,while some experiments will be on real online communities.With this regard,we are adding weblog functionalities to a classical music recommend system:CoCoA1[5].In this case,our goal is to have a system running with real users(about1400daily now) that can express what are the bloggers they trust and what are the classical tracks they like in which we will be able to test our proposed solution against a commu-nity of real users.The structure of the proposal is the following.In section2,we present the past work relevant for our research line.The chosen subject is quite interdisci-plinary and so we refer to different topics such as Rec-ommender Systems(section2.1),Reputation-aware sys-tems and Trust metrics(section2.2),Peer-to-Peer and distributed architecture(section2.3),and Weblogs and Semantic Web(section 2.4)Section3provides the motivating context,i.e.the reasons we claim make senseful and innovative to ex-ploit trust information with the goal to create better Recommender Systems.In section4we outline what are the research prob-lems arised in the new context while in section5we clearly indicate the research problems we address along with our proposed solution.We present the experiments we plan to run in order to evaluate our proposed solution in section6.Finally in section7we present the work already done and a detailed road map for what must be done,while in section8we indicate what we leave as future work.2State of the ArtIn this section,we present the past work relevant for our research line.The chosen subject is quite interdis-ciplinary and so relevant papers come from many dif-ferentfields.First,we analyze Recommender Systems, automated tools that promise to give a solution to in-formation overload.Then we concentrate on systems that take explicitly into account trust and reputation, analyzing how this information is currently used.We then present peer-to-peer and distributed architectures because it is in this environment that trust information can really unveil its power being every peer free to be-have as it st,we analyze weblogs that are a recent and very interesting Internet trend.They can be seen as thefirst,real instantiation of the SemanticRecommender Systems for online tests.An interesting application for RS is in the music do-main:we developed a RS for suggesting classical music compilations[5].The running application can be found at http://cocoa.itc.it.We also proposed recommendation strategies to adapt Internet radio programs on thefly[6].Interesting possibilities still unexploited for recom-mender algorithms come from social network analysis by which it is possible to spot out what are the neighbours of an user by her actions.Location-aware computing also offer new frontiers for RS[9],in fact,information about the current location of the user opens new interesting possibilities for recommendation algorithms.Lately,there have been some attempts to move CF in a distributed environment.John Canny in“Collabora-tive Filtering with privacy”[10]criticizes the centralized approach in which all the user data resides on a cen-tral server and proposes an alternative model in which users control all their log data.He also describes an al-gorithm whereby every single user of the community can compute a public“aggregate”of their data that does not expose individual user’s data.Then every single peer can use this aggregate to compute recommenda-tions.The proposed algorithm uses factor analysis(a technique of dimensionality reduction similar to singu-lar value decomposition but more efficient)and crypto-graphic techniques(asymmetric encryption).The key idea is homomorphic encryption to allows sums of en-crypted vectors to be computed and decrypted without exposing individual data.With this regard,another interesting papers is“Col-laborative Filtering In A Distributed Environment: An Agent-Based Approach”[19].The authors have proposed a system called iOwl to exchange meta-data about web surfing activity between peers.They have also created a usable application downloadable at .Essentially,a modified browser records user’s clickstreams and data mining techniques extract profile data,such as usual navigation patterns. This metadata are exchanged with other peers and are used to self compute recommendations of possibly inter-esting URLs.2.2Reputation-aware systems and trustmetricsNowadays,with the emergence of online communities,e-marketplaces,weblogs and peer-to-peer communities,a new kind of information is available:rating expressed by an user on another user.We call this information trust.Moreover,there are examples of communities that are no more only virtual and online but have started to produce important effects in the real world:for ex-ample, is the cause of a big movement of goods and money in the real world even if it’s noth-ing but a virtual community.Other examples of no-more-so-virtual communities are weblogs(the blogo-sphere),auction sites(,yahoo auction,ama-zon auction and many more),,aff, ,peer-to-peer networks(edonkey network,reputation-aware gnutella network).In the future with the spread of mobile,pervasive and ubiquituous comput-ing this information will become even more available and useful and concrete.We can ask ourselves“Is this information useful?”.The answer is a great yes.For instance,[36]a site devoted to online auctions,allows traders to rate their partners after a commercial transaction;in this way,it can compute and show to users the reputation as a seller and buyer of every member and this simple number near the usernames is what enables big economic movements to happen in the real world between people geographically distant,who never met before and who will never meet in future.This is the main reason for the great(economic)success of this dotcom.Other examples are ,the“news for nerds”site where everyone can post news story and rate other users depending on posts they submit;here reputa-tion is used to keep noise-to-signal ratio low and to give special emphasis to interesting posts;aff is a sys-tem that exploits trust elicitation in order to democrati-cally and distributedly decide which open source projects are the more promising for the community and worth funding.There are some attempts to build reputation-aware systems also on top of current P2P networks:on eDonkey network with the open source client eMule ()and in reputation-aware Gnutella servents[11].Trust has been introduced with the goal of blocking and spotting out leeches(peers who only download without sharing)and misrupters(peers who pollute the network with faked items).[35]provides a complete analysis of most of the cur-rent existing reputation systems.Trust is not a strange or artificial concept.If you are member of a mailing list,surely you have experienced the following:the topic of the mailing list is of inter-ested for you but while you are eager of reading mails from some members(you value/trust them a lot!),you don’t want to be bothered by mails of some other mem-bers/spammers/disrupters(you totally distrust them!).Tools that can help you in saying how much you trust any member will enable services personalizing for you the access to relevant information minimazing the quantity of bad quality content you have to process.The same can be said for newspapers,journalists,politicians,mu-sicians,and people in general.There are many different definitions of trust and re-lated concepts such as reputation and reciprocity.This shouldn’t surprise since trust is a very human and social concept and it is in the speculation of men since thefirst philosophers.Moreover these concepts have been stud-ied by researchers and thinkers in differentfields other than computer science such as economics(game theory), scientometrics(or bibliometrics),evolutionary biology, anthrolopoly,philosophy and sociology[31].We will clearly state our definition of trust and re-lated concepts in section5.2.For now,we just point out how,in computer science literature,trust is intended mainly as a diadic quantity 3involving two peers:“a subjective expectation an agent has about another’s future behaviour”[31]while repu-tation is mainly seen as property of a peer assigned to her by the embedded social network and computed from the many trust relationships:“reputation is the memory and summary of behaviour from past transactions”[32] (chapter17).Anyway they can be seen as two sides of the same concept and are often used as sinonyms.Many persons believe the world of future will be based on reputation:reputation will become the only “currency”of our life[37].Also one of the sci-fi2003 bestsellers explores the topic:in“Down and out in the magic kingdom”[12],Cory Doctorow envisions a near-future realistic world that’s seen“the death of scarcity”, where nanotechnology takes care of everyone’s basic needs and there is no more need for money.Instead, what the population aspire to is“Whuffie”,a sort of reputation capital representing the approval and respect of your peers.Marsh[28]was thefirst to introduce with his PhD thesis a computational model for trust in the dis-tributed artificial intelligence community and there are some attempts to scientifically understand what trust and reputation concepts can represents for computer sci-ence[1,45,23,22,2,31]but it is worth noting how the research is very recent and how most of the proposed rep-utation systems for online communities have been con-structed with an intuitive approach and without a formal model and a deep learning from approaches of the social sciences.Often these papers just provide some possible, reasonable intuitions but without solid ground or moti-vations;sometimes there aren’t even tests on simulated data and in general no evaluation with real communi-ties.Researching in this direction is indeed very needed in order to discover potentialities and possible problems and to design systems that can get full advantage of this information.It is of course not possible to build a direct reputa-tion relationship with every other peer,so it is impor-tant to share judgements about other peers.Sen et al. in[41]demonstrate that cooperating agents who share their opinions about other agents perform better(i.e. maximize individual utility)than selfish ones who don’t collaborate.Some trust metrics have been proposed:among oth-ers,Advogato[26]and Fionna[25].It is important to underline how all these metrics are simple and intuitive and that more advanced ones can be built only evaluating them in real,used systems.With this regard,the more interesting projects are:NewsMon-ster and BlogNet2,two OpenPrivacy projects working with weblogs and news channels.Many of these systems compute a global reputation value for every single peer,with this regard they are very similar to PageRank[33],the algorithm used in for deciding the importance of an Internetpage.We will see some weaknesses of this approach in section3.3Decentralized Meta-Data Strategies at /Decentralized Strategies-neat.html4 and 4phenomena5:everyone tries to increase her utility from them,so overuses them even without a real need and this leads shortly to total consumption and unavailability for everyone.The other approach instead says that if the effort in contributing to create a resource(an up to date repository of informations,for example)is very low,than new resources can be created by the spontaneous work of all the users of a system and made available to everyone. Moreover if the resource is not a consumable one(as it is the case with bits)there is only increase of available resources.In the following we will use the term P2P in a more relaxed ethimological sense:we will consider a peer ev-ery logical indipendent entity that can be identified in a unique way and is able to expose some information.To be clear,we’ll consider a weblog(see subsection2.4)as a logical peer.How P2P architectures are relevant for our research? We have seen that in a decentralized environment ev-ery peer is free to behave as she prefers.Anyway one important way to incentive good behaviours is the pos-sibility for peers to be rewarded or punished,essentially to enable the explicitation of trust(see section2.2).In systems like ,every peer is free to behave badly,for example to not send the item she has already received the payment for.Instead peers almost always behave correcly because the opinions of other peers con-tribute to create their reputation in the system and rep-utation is a value because it allows to make trade and to get better prices.The same can be seen on where having a good reputation is equal to be recognized as a guru by the community or in aff where receiving many trust statements means you are valuable for the open source community(i.e.a skilled hacker)and this means you couldfind a very interesting job easily.We can say that in a decentralized environ-ment that takes into account trust,there are incentives to well behave.This incentives allows to“harness the power of disruptive technologies”[32]such as peer-to-peer systems lowering down the inherent risks of these distributed environment.2.4Weblogs and Semantic WebA very interesting phenomena of the last years in Inter-net are weblogs(often contracted in blogs).They are a sort of online diary,a frequently updated web site arranged chronologically,very easy to create and maintain that doesn’t require knowing HTML or programming.It provides a very low barrier entry for personal web publishing and so many millions of people in the world maintain their own blog and post on it daily thoughts.Their relevance is confirmed by the following facts: in February2003,Google bought Pyra Labs,a company that created some of the earliest technology for writing weblogs and its website,;Stanford and Har-vard are promoting their use among their students as a valuable mean of publishing of research ideas and results.6/rss7/foaf/0.1/8/smbmetaintro.html95but an extension of the current one,in which infor-mation is given well-defined meaning,better enabling computers and people to work in cooperation”(from /2001/sw/).It is everything about converging on some XML stan-dard formats and succeeding in getting them widely used.Nowadays,weblogs seems the only possible walk-able way in this direction.Weblogs are important for our research line because they can be the empowering tool fostering the availabil-ity of an always up to date distributed datasets of cor-relations among peers(trust)and with items(ratings). We will see in section3how one factor that slowed down research in RS was the unavailability of public datasets and testbeds.Weblogs promise tofill this gap.3Motivations and impactWe have already seen the great potential behind Recom-mender Systems(RS).In particular we have examined the simple but very effective intuition at the base of Col-laborative Filtering(CF)that is automating the“word of mouth”process.Anyway RSs(especially based on CF)have some weaknesses we will examine in a short.Essentially,we can say that CF automates the pro-cess too much forgetting the social dimension of the en-vironment.It doesn’t take into account what are the opinions of people about other people but it simply tries to predict them based on similarity about how they rate items.We claim that taking into account direct judge-ment of users on users(trust)can enhance RSs perfor-mances.An analysis of the concept of trust and its use has already been done in section2.2.We have outlined how many current online communities make use of the con-cept of trust in order to have a decentralized control over the system.It is worth noting again how many of these systems compute a global value of trust for every peer (for instance, shows shows near every username a star of different colors where a color means a certain level of reputation in the community computed on the basis of every positive and negative feedback given by users after every transaction).We claim this is a non-sense because an user can be valuable for one peer and not for another peer.We claim only personalized and subjective computation of reputation can be really use-ful and not attackable.We use the term“indirect trust”to indicate a predicted trust to emphasize how this value has only meaning as a relation among two peers and not as a global value equal for everyone.In the following we will continue using the term “peer”intending a logical indipendent entity that can be identified in a unique way and is able to expose some information fetchable by other peers.Let come back to our starting point:Recommender Systems weaknesses;in the following we will examine them one by one explaining also how trust-awareness can help in overcoming them.Content-based RSs require human editors to tag andclassify items.Content-based RSs tries to suggest to the user items similar to the ones she liked in the past.To do this,they need a representation of the items in term of features.This representation can be extracted automatically from items whenever it is possible:for example,for read news,it is possible to extract the words an user has interested in simply by parsing the text.Even if this methods often doesn’t capture the essence of the tastes of an user,they are used because very easy to implement.Content-based RSs have more problems when items cannot be parsed by machines:for example, it is impossible or very difficult nowadays for a machine to extract meaningful features such as genre,authors from a song or a moviefile.In this case we need human beings to tag and classify items.This introduces a lot of problems:first,it is not easy to decide which are the right features we have to tag(genre,instruments,year, ...)and tagging is an expensive,boring,error-prone, subjective activity.Moreover for some items such as jokes is almost impossible tofind what the right features are.Our approach is to use only information provided by the user:ratings on items and trusts on users.In this way our resulting system is totally independent of the specific items features that are not taken into account and can be applied to whatever domain.Confidence in computed User Similarity is often very low.The information expressed by user about who and how much she trusts(“PeerA trusts PeerB0.9”)is con-sidered to be correct information,with total confidence.Instead the information“peerA is similar to peerB0.8”is computed by the system.The confidence in the correct-ness of this value depends on the quantity of information that is available and the algorithm that is used.We will define better confidence in section5.2,for now we use it for indicate the reliability of a computed value,the degree of certainty the system has in it.If we consider every user as a vector of ratings on items and we put them in a matrix,the resulting ma-trix users×items,the traditional input of Collaborative Filtering techniques,is very very sparse(for example, Eachmovie[29]dataset is97.4%sparse).This is totally normal,infact,if we need something tofilter out bad items or to suggest interesting ones,this means there are many items we are not going to try and,consequently,to rate.This results in our“profile”consisting of ratings of few items over a huge set.This is an inherent weakness of systems relying only on users rating items and this is even esacerbated in the early phases of RS life cycle when really few ratings are available.In general,sparseness means that overlapping of pro-files of2users(i.e.the number of items they both rated) is very low and their user similarity often not computable or anyway computable with low confidence.Singular Value Decomposition has been proposed as a solution for dimensionality reduction and consequent reduction of sparseness[38]but it is not clear if when you start with a so sparse matrix,it is really the amount of in-formation available that is too less.Another proposed solution are hybrid systems[40,17],in which we have to 6rely also on features based on content,especially in an initial phase.Another problem resulting in low confidence in com-puted User Similarity is the so called cold start problem. It arises when a new user enters in a new domain where, clearly,she has expressed no ratings.In this case,CF cannot compute similarity andfind neighbours and com-pute recommendations.It is even possible she is not able or doesn’t want to rates items in this domain(let’s take for example a computer scientist wanting to start reading “black holes”physics books and knowing nothing about them).However,in a social setting(as Internet at broad is),the“new”user probably has(or can easilyfind)some trustworthy,valuable peers;then these explicitly stated valued neighbours will guide the recommendation pro-cess that she has always under control with the possibil-ity of expressing trust on new discovered peers,to rate recommended items,to calibrate her own tastes and to change every previously expressed statement.Summarizing,we claim that expressing a trusted peer is easier and more natural for users and allows an easy bootstrap of the system;direct trust information has to-tal confidence and is more reliable that computed user similarity.Moreover trust has some inherent transitivity properties(if A trusts B and B trusts C,A will probably trust C).Considering the“Six Degrees of Separation”folklore axiom10,we can claim that a small number of trust statements can easily cover a big portion of the social network if we exploit trust transitivity and propa-gate trust.In the case of an undirect trustable peer,the confidence is no more total but it is anyway a very useful information in order to reduce significantly sparseness.There can be the case when user similarity and trust contradict each other;we will try to understand better what this means in the following,anyway taking into account also trust should be better than just relying on user similarity.CF techniques are attackable by malicious users.CF takes into account every peer in the same way. In this sense it has no means at all to discover malicious peers.For this reason,malicious peers that knows how the algorithm works can easily exploit it in order to in-fluence the created recommendations.For example,let suppose a malicious user wants the RS to recommend ItemSpam to PeerA:she can create FakePeer,copy the profile of PeerA and add a good rating to ItemSpam.In this way the RS willfind FakePeer as the most similar to PeerA and recommend ItemSpam to it.We think it is very important to take into account attacks in a distributed system where every peer is free to behave as it prefers.There is some research about this in P2P systems[26,22]but we know of no papers pointing out this problem for RSs.Moreover in a distributed virtual systems,the fact that it is easy to create infinite new identities means that you cannot trust a peer unless you have grown a positive(direct or indirect)trust in her[14].。

相关文档
最新文档