Knowledge Network System (KNS) by Evolutionary Collective Intelligence (ECI): Model, Algorithm and Applications
Tao Xiang, Ziliang Huang, Peng Bai, Congrui Ji, and Zhiyong Liu**
Visva, Inc. 20410 Town Center Lane, Suite 295 Cupertino, CA 95014, U.S.A.
{taoxiang, zilianghuang, pengbai, congruiji, zhiyongliu@visva.com }
Abstract. Aiming at overcoming some inherent drawbacks and bot- tlenecks encountered by the conventional Knowledge, Recommendation, Search, and Social Systems, in this article we introduce the Knowl- edge Network System (KNS), a novel type of knowledge graph which is constructed by a new proposed algorithm, the Evolutionary Collec- tive Intelligence (ECI) algorithm. The ECI, an agent-machine interac- tive algorithm, constructs the KNS by iteratively recommending inter- esting/matched samples/files to the agents, and meanwhile taking ad- vantages of the collective intelligence of the agents. The ECI based KNS, to the best of our knowledge, is the first attempt in literature that in- tegrates the functions of knowledge network construction, high-quality recommendation, new types of search and social in a same framework. Some real and potential applications of KNS and ECI are discussed, and a real system named VISVA is provided to demonstrate their efficacy. Some open problems for future works are also summarized in the end.
Keywords: Knowledge Network System, Evolutionary Collective Intel- ligence, Knowledge System, Recommendation System, Search Engine, Social Network System, Internet of Things, VISVA
Introduction
Nowadays, people query/search/receive information from the Internet by using a variety of tools, Knowledge System such as Wikipedia, Search Engine such as Google, Recommend Systems such as TopBUZZ, and Social Network System (SNS) such as Twitter, to name a few. The underlying technologies behind them have proven quite effective during the last decade. Generally, the Knowledge System employs entry to facilitate the query, Search Engines make use of key words matching, Recommend System works by mining the correlations between the articles and/or users, and SNS relies on the follow relationship between users. Though these tools have achieved great success by now, all of them exhibit some inherent drawbacks.
** Corresponding author
The entry system makes query convenient to implement, but it somehow suf- fers from at least the two facts. First, the entry system is tedious to maintain, especially in nowadays many new ideas and concepts emerge almost everyday; second, it is hard to precisely locate in the entry system some multidisciplinary knowledge, let alone those concepts beyond words and unnameable. For the key words based search engines, only the articles containing the key words can be found, but cannot hit those closely related materials but without explicitly containing the key words; it is even harder, if not impossible, to recover those nonverbal medias such as image, video and audio files. Huge number of rec- ommendation algorithms have been proposed in recent years. The content or user based recommendation recommends the content based on some similarity between some related contents or users, but hard to generalize to find other po- tential interesting items; the knowledge graph based recommendation can work in a more systematic way, but how to formulate the knowledge graph (especially suitable for recommendation) remains to be a tough task. The follow relation- ship in SNS means that the follower receives all of the information posted by the one followed, however, in general not all but only a part of the posts might be interesting to the follower, consequently making the follower receive much noise. In view of the above considerations, in this paper we introduce the Knowl- edge Network System (KNS), which is gradually formulated by the Evolutionary Collective Intelligence (ECI), a novel agent(human)-machine interactive learn- ing algorithm proposed particularly for the KNS. The main contribution of this
article is summarized as twofold:
- To overcome the drawbacks of the knowledge and recommendation systems etc., we propose the ECI algorithm to construct the KNS, a new type of knowledge construction, storage and distribution system.
- A real application based on the KNS and ECI is presented to illustrate their efficacy.
The remainder of this article is organized as follows. The KNS and ECI algorithms are introduced in Sections 2 and 3, respectively, followed by some theoretical analysis and discussions in Section 4. After giving some real and po- tential applications in Section 5, Section 6 concludes this article by summarizing some future works.
Knowledge Network Systems
Some notations used in the KNS and ECI algorithm are firstly given as follows.
-
- Knowledge file, denoted by f , is the basic input or sample of the knowledge system. Typical types of knowledge file include such as message, article, image, video, and audio file.
- Knowledge item, denoted by v, exactly a knowledge node in the KNS, rep- resents one basic knowledge point in the KNS. Each item consists of one or more knowledge files, or in other word, the knowledge item is a collection of
these files, which share some common characteristics. Since as will see below each node in the KNS is also accompanied with some agents, sometimes we may also call each item a group
剩余内容已隐藏,支付完成后下载完整资料
基于演化群智的知识网络系统: 模型,算法和应用
作者:Tao Xiang, Zi-Liang Huang, and Zhi-Yong Liu
摘要
本文针对传统知识、推荐、搜索、与社交等系统面临的缺陷 和瓶颈问题, 引入一种知识网络系统(Knowledge Network Sys- tem, KNS), 其本质是一种知识图, 通过一种新提出的演化群智
(Evolutional Collective Intelligence,ECI)方法来构建。ECI是一种智能体(人)-机器交互算法,利用智能体的群体智慧(或称群智)迭代的向智能体推送匹配的样本、文件等来构造KNS。据文献调研,基于ECI的KNS首次尝试将知识图谱、高质量推荐、新型搜索和社交网络等功能整合在同一框架内。本文主要讨论KNS和ECI的一些已有和潜在应用,并介绍名为VISVA的实际系统来证明其有效性,最后总结讨论一些值得进一步研究的开放问题。
关键词: 知识网络系统,演化群智,知识图谱系统,推荐系统,搜索引擎,社交网络系统,物联网,VISVA
引言
如今,人们通过知识系统(如维基百科)、搜索引擎(如谷歌)、推荐系 统(如今日头条)、以及社交网络系统(推特)等工具从互联网上查询、 搜索、接收信息。过去十年里,这些应用实践充分证明了其所采用的基础
技术的有效性,比如知识系统采用词条来方便查询,搜索引擎利用关键词 匹配,推荐系统利用文章或用户之间的相关性,社交网络系统依赖用户之 间的互相关注关系。尽管这些基础技术已取得巨大成功,但仍显示出一些 固有缺陷与瓶颈问题。
虽然基于词条的知识系统使得查询易于实现,但是它至少受到两种不 利影响:一是,在当前新思想、新概念不断涌现的情况下词条库本身维护 更加繁琐;二是在词条库有时很难准确定位一些交叉学科知识和语言难以 描述的抽象概念。基于关键词的搜索引擎一般只能找到包含该关键词的文 章,但是难以找到不直接包含该关键词但内容密切相关的资料,尤其是图 像、视频和音频等非文字资料。近年来涌现出非常多的推荐系统:基于内 容或用户的推荐系统利用内容间或用户间人为定义的相似性来推荐内容, 难以泛化到其他潜在相关的内容;基于知识图谱的推荐系统构建有效的, 尤其是适合推荐任务的知识图谱,尤其是仍是一项艰巨的工作。社交网络 系统的关注机制使得“粉丝”被动接受发布者发布的所有信息,但事实上 “粉丝”会接收到很多噪声信息,他所感兴趣只是一部分内容而已。
有鉴于此,本文引入一种知识网络系统(Knowledge Network System, KNS),并针对性的提出用于KNS的智能体(人)-机器交互学习算法,即演化群智(Evolutional Collective Intelligence,ECI)算法。本文的主要贡献为以下两个方面:
-
- 针对知识图谱和推荐系统的不足,提出ECI算法构建KNS,以新的方式实现知识建构、存储和分发。
- 通过实际应用实例,证明KNS和ECI的有效性。
本文分为以下几个部分: 第2节与第3节分别介绍KNS和ECI算法, 第4节进行相关的理论分析与讨论, 第5节介绍一些已有和潜在应用, 第6节总结本文并讨论下一步工作。
知识网络系统
首先介绍KNS和ECI算法使用的概念与符号。
-
- 知识文件(Knowledge file) 指知识系统的基本输入或样本, 以符号f 表示。典型的知识文件类型有消息、文章、图像、视频和音频等。
- 知识项(Item),即知识网络的节点,表示KNS的基本知识点,以符号v表示。单个知识项由一个或多个知识文件支撑组成,也可以说, 知识项是具有某种共同特征的知识文件的集合。此外,KNS中每个节点与若干个智能体相关联,从聚合智能体成组的角度来看,知识项也可称为群组。知识项、图节点、和群组三个术语在本文通
- 智能体(Agent),指对某些特定知识文件感兴趣并可以做出反馈的实体,可以是自然人或其他智能实体,以符号a表示。如果智能体a对推送给它的文件f 感兴趣,我们认为f 与a相匹配,此时f 对a来说是有意义的信息,反之,则为噪声。基于此,推送给智能体a的文件集合 可以按下式定义信噪比(signal-noise ratio, SNR):
#匹配文件
snra = #未匹配文件 (1)
知识网络系统KNS记为K,可以定义为动态有向知识图K = {V, E},其中V 表示图节点集合,E = (vx, vy)是图边(节点vx指向节点vy的有向边) 集合,如图1所示。上述定义中,动态意味着包括V 和E在内的图结构可能随着时间演化。如下文所述,K可以从一个不包含任何节点的空图演变而 来,演变完成后也可以是具有固定节点和结构的静态图。
KNS中的每个节点v表示一个知识项,有两层含义,一是包含的知识文 件,二是属于它的智能体。所以v可以表示为
v = {fv, av}, fv = {fv, fv, ...fv}, av = {av, av, ..., av } (2)
1
2
k
1
2
n
要说明的每个智能体可能与很多知识项相关联,因此知识项v的构建本质 上取决于fv而非av。此外,每个v中收集文件fv的方式是通过智能体投票产生,这与人工设计生成条目的方式截然不同。
基于fv,给定的两个节点vx与vy,两个有向图边exy与eyx的权重值可以由下式分别计算
| cap; |
| cap; |
exy =
fvx fvy
|fvy | , eyx =
fvx fvy
|fvx| . (3)
图 1: 知识网络系统KNS示意图
两个节点之间的每一条有向加权边不仅度量节点间相似性,同时也度 量其层次关系。例如,给定一对有向边exy = 0.9与eyx = 0.5,两个相关节点vx与vy间的相似性可以直接由1 (exy eyx) = 0.7计算得到,同时不等式exy gt; eyx意味着vx在K相对来说占据更高的层次等级。进一步说,一旦建立KNS我们还可以使用一些分层谱聚类工具[6, 5]来挖掘其分层知识结构。
2
在实际应用中,我们可以为加权边设置一个阈值Te,如果exy lt; Te则将exy置零,这样KNS一般为稀疏图,意味着只有很小一部分节点对被激活。从存储和计算的复杂性来说,这种稀疏性对于大规模的K尤为重要。
演化群智算法
本节介绍基于智能体(人)-机器交互专门用于构建KNS的演化群智算法
(ECI)。
ECI算法利用智能体和机器间的交互来实现:机器向智能体推送知识文 件,智能体选择是否激活推送文件。智能体和KNS间的交互关系如图2所 示,图中上层和下层分别代表智能体与KNS,一个智能体连接到KNS的某 些节点,表示该智能体与相应知识项匹配或者说属于相应群组,ECI算法的 设计正是利用智能体和KNS间的这种关系。ECI算法包含两个目标:第一是生成的KNS中每个知识项中的多个文件应包含共有知识,第二是可实现高
图 2: 智能体与KNS交互示意图
质量推送,即高信噪比的推送。
ECI算法基本流程图如图3所示,首先由系统或智能体发布新文件f , 然后f 以随机或其他更复杂的方式推送给智能体acirc;( 流程图中的初始传播:initial spread ), 比初始传播外更重要的是,f 还被推送给与激活的知识项相关联的部分或全部智能体(流程图中的KNS传播:spread by KNS ),然后评估文件f 是否符合加入知识项circ;i的标准(流程图中的并入知识项:add into the item),同时评估其是否适合作为创建新知识项的创始知识文件(流程图中的新知识项:new item)。
图 3: ECI算法基本流程图
KNS传播(spread by KNS ) 是ECI算法的核心步骤, 对实现上述构建KNS与高信噪比推送的目标至关重要,下面给出详细介绍。文件f 通过初始传播(initial spread )与KNS传播(spread by KNS )两种方式推送给智能体。初始传播用来为新文件随机选择智能体,其信噪比基本等同于随机 推送信噪比,因此主要依靠KNS传播提高系统整体推送信噪比,通过投票 机制和多轮传播机制来实现:投票机制指只有与一个知识项关联的大多数 智能体都应能与一个文件匹配,或者说大多数智能体都为该文件投票(投 票比例应大于预设阈值),那么这个文件才能在该知识项中的传播;多轮 传播机制指每个文件在一个群组中要逐轮传播,每一轮中只被推送给部分 智能体,只有在该轮中通过投票表决的文件才会进入下一轮传播。投票机 制保证只有通过智能体投票的文件才可能在相关群组内传播,多轮投票机 制通过逐步渐进的抽样机制保证知识文件的高信噪比。
对于其他模块,新知识项(New item)指如果文件f 的智能体匹配数大于设定的阈值,则以f 作为创始知识文件创建一个新知识项v,并将与之匹配的智能体加入该新知识项,在一个知识项或者说群组中通过所有多轮传 播投票考核并满足一些其他约束1的文件将被添加到该群组。
理论分析与讨论
基于上文介绍的KNS和ECI,下面给出一些理论分析与讨论。
关于智能体
检查智能体a与文件f 是否匹配是ECI算法的一个关键步骤,对于自然人智能体,获得文件推送后,匹配关系由点击、评论、保存等用户操作来实 现,而对于非人智能体就需要在其与文件间引入额外信息来实现匹配操 作。下面针对非人智能体我们引入知识单元u 的额外变量,虽然自然人智能体不需要知识单元的概念,但作为隐藏变量它也有助于从理论上分 析KNS和ECI。
知识单元u是知识网络系统中不能分解的、最基本的知识元素,一个知
1例如在添加f 到v之前,应检查f 与知识项v 中已存在文件之间的契合度。
识文件f 可以作为一个或多个知识单元的集合,从智能体角度看知识单位可以视为兴趣点,也就是说智能体a也可以视为知识单元或称兴趣点的集合, 所以给定一个知识单元的集合
U = {u1, u2, ...., um}, (4)
知识文件lt;
剩余内容已隐藏,支付完成后下载完整资料
资料编号:[254338],资料为PDF文档或Word文档,PDF文档可免费转换为Word
以上是毕业论文外文翻译,课题毕业论文、任务书、文献综述、开题报告、程序设计、图纸设计等资料可联系客服协助查找。