基于复杂网络的推荐算法和合作行为研究

发布时间：2018-04-03 12:37

本文选题：复杂网络　切入点：演化博弈　出处：《中国科学技术大学》2011年博士论文

【摘要】：随着社会和经济的快速发展,人类的生活方式发生了巨大的变化。对人类行为的分析与解读已成为一个重要的研究课题。现实社会中有许许多多的复杂系统,这些复杂系统可以抽象为各种复杂网络。近些年来,复杂网络的兴起及快速发展,使其已经成为探索自然界和社会系统中很多问题的重要工具。另外,计算机技术的快速发展,使大量统计数据的采集和建模模拟都成为可能,这也为很多问题的研究提供了新的思路和方法。本文中我们首先介绍了一些复杂网络的背景知识,然后重点介绍了基于网络的推荐算法和网络上合作行为研究。随着人们面临的信息爆炸性增长,在这些海量信息中想找出自己感兴趣的信息越来越难。个性化推荐被认为是解决这个问题的有效途径。个性化推荐系统是通过对用户历史行为进行分析,揭示用户习惯和喜好。然后建立相应的推荐算法,对每个用户给出一个推荐列表使他可以快速地找到自己感兴趣的信息。我们的主要工作是基于网络推荐算法的改进。不同于以往的算法,我们考虑接收节点所拥有资源的影响,即物质扩散资源的相对的量要比绝对量更具有参考意义。文中引入一个可调参数来研究资源接收者对接收资源的响应与其自身度的相关性。利用ranking score和另外两种测量个性化程度的评价方法,模拟发现适当的加强对度大节点(即流行产品)的推荐,不仅可以大幅提高算法的精度,而且还可以大幅提高算法的个性化程度。另外,我们还提出了一种改进的协同过滤算法。与原来的余弦相似性计算方法不同的是,在计算事物相似性的时候不仅考虑他们公共属性的影响,还考虑公共属性流行程度的影响。我们认为公共属性越流行的,其对相似性的贡献越小,越不流行的属性,才具有更个性化的特征,其对相似性的贡献也就越大。引入这种公共属性对事物相似性的负相关作用,我们提出了修正的协同过滤算法(MCF)。模拟发现不论是以用户为基还是以产品为基,新算法所得推荐列表的准确性都比原来的算法有很大的提高。我们知道很多系统都可以抽象成特定的网络,且很多时候网络上个体的相互作用我们可以抽象为个体的博弈过程。通过研究个体间的博弈过程,可以对自然社会中合作现象有一些探索。在现实生活中,普遍存在着各种不确定的因素,这些不确定因素可能会直接影响到个体的适应能力。因此个体的适应能力会存在一定程度的涨落。在本文中我们以囚徒困境博弈为例,研究了个体适应性的涨落在二维规则格子上和NW小世界网络上的合作演化情况。引入了两个参数α和β表征适应性涨落的情况,其中α表示适应性涨落的个体比例,β表示适应性涨落的幅度。模拟发现在涨落幅度不是很大而涨落人数适中的情况下,合作频率会得到加强。我们的结果对理解现实生活中涨落情况下合作的涌现有一定的帮助。另外我们还研究了遗传性对公共物品博弈和囚徒困境博弈的影响。遗传性在真实的生物系统和社会系统中是一种普遍特性,子代可以遗传父代的相关特征,例如适应性策略以及父代学习策略的方式等。考虑个体适应性受遗传因素和当前收益影响,研究发现个体遗传性的引入,可以使合作簇有效的抵御背叛者的入侵,从而使合作行为能够持续存在。最后我们还研究了地理距离对命名博弈的影响。众所周知,人类语言的形成是个复杂而漫长的过程,各地的语言有各地的特点,但是随着社会的发展,不同语言之间相互影响和渗透,其中地域在语言的形成和发展过程中的影响不可忽视。我们以一个简单的命名博弈为模型,研究地理距离对整个博弈过程的影响。模拟发现过长或过短的加边长度都不利于最终一致的形成,只有适当的加边长度才有利于一致意见的形成。文中还详细研究了整个博弈过程中相关参数的变化,例如最大词汇数目不同词汇数收敛时间等。
[Abstract]:With the rapid development of society and economy, great changes have taken place in people's way of life. The analysis and understanding of human behavior has become an important research topic. There are many complex systems in real society, these complex systems can be abstracted into a variety of complex networks. In recent years, the rapid development of complex network and the development, which has become an important tool to explore many problems of natural and social systems. In addition, the rapid development of computer technology, the acquisition and modeling simulation of a large number of statistical data are possible, it also provides new ideas and methods for the study of many problems. In this paper, we firstly introduce some complex networks the background knowledge, and then focuses on the research of cooperative behavior and recommendation algorithms based on the network.
With the explosive growth of information people face, find their interested information more and more difficult to think in the vast amounts of information. The personalized recommendation is considered to be an effective way to solve this problem. The personalized recommendation system is carried out through the analysis of user behavior, revealing user habits and preferences. Then establish corresponding recommendation algorithm for each. The user gives a recommendation list so that he can quickly find the information they are interested in. Our main work is to improve the recommendation algorithm based on network. Unlike previous algorithms, we consider the impact of node resources, relative to the amount of material resources to diffusion is more important than the absolute amount of introduction. An adjustable parameter to study the correlation between the response of the resources receiver and his degree in this paper. By using ranking score and other two kinds of personality measurement The evaluation method of simulation, find the appropriate strengthening of the degree of nodes (i.e., popular products) recommended, not only can greatly improve the accuracy of the algorithm, but also can greatly improve the degree of personalization algorithm. In addition, we also propose an improved collaborative filtering algorithm. With the original cosine similarity calculation is different the method is in the calculation of the similarity of things not only when they consider the impact of public property, but also considering the influence of public property. We believe that the popularity of public property more popular, the similarity contribution is small, do not attribute the more popular, it has the characteristics of more personalized, contribution to their similarity have more. The introduction of this public property similar negative effect of things, we propose a modified collaborative filtering (MCF). Simulation shows that as a user base or in product based, new algorithm The accuracy of the recommended list is much better than that of the original algorithm.
We know that many systems can be abstracted into a specific network, and a lot of time on the network of individual interactions we can be abstracted as the game process of individual. By studying the game process between individuals, there can be some exploration on cooperation in natural and social phenomena. In real life, widely exists in various uncertain factors, these the uncertain factors may directly affect the individual's ability to adapt to fluctuations. Therefore the individualsability there will be a certain degree. In this paper we use the prisoner's dilemma game as an example, the evolution of cooperation study fluctuation individual adaptability in two-dimensional regular lattice and NW small world network. The introduction of the two parameters the characterization of adaptability and beta fluctuation, the proportion of individuals said alpha fluctuation adaptability, beta fluctuation amplitude. Said adaptive simulation found in fluctuation is not very big rise Drop the number of moderate conditions, the frequency of cooperation will be strengthened. Our results on the fluctuations in real life under the condition of understanding the emergence of cooperation will help. In addition, we also investigated the effect of heredity on public goods game and the prisoner's dilemma game. Heredity is a common feature in biological systems and real in the social system, the offspring can be related to genetic characteristics, such as the parent and parent adaptability? Strategy learning strategies etc. the individual adaptability by genetic factors and the effects of revenue considerations, research found that the introduction of individual genetic, can make clusters effectively resist the intrusion of defectors. Thus, cooperative behavior can be sustained finally. We also study the effect of geographic distance on the naming game. As everyone knows, the formation of human language is a complex and lengthy process, features around the language around, But with the development of society, between different languages, mutual influence and infiltration, which affect the region in the process of formation and development of language can not be ignored. We use a simple naming game model, effect of geographical distance on the whole game process. The simulation is too long or too short length are not conducive to the border a final agreement form, only the appropriate length is bordered to form a consensus. The changes also discussed in the relevant parameters of the whole process of the game, such as the maximum number of different vocabulary vocabulary?? the convergence time.

【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2011
【分类号】：F062.6;F224.32

【相似文献】