Gotowa bibliografia na temat „Complex Social Networks”

Orientadores: André Franceschi de Angelis, Regina Lúcia de Oliveira Moraes<br>Dissertação (mestrado) - Universidade Estadual de Campinas, Faculdade de Tecnologia<br>Made available in DSpace on 2018-08-27T04:53:50Z (GMT). No. of bitstreams: 1 Libardi_PaulaLucieneOliveira_M.pdf: 1610224 bytes, checksum: a08b75cd1a30c421927617ee8b6ac8d4 (MD5) Previous issue date: 2015<br>Resumo: A identificação de usuários falecidos em Redes Sociais Online é um desafio em aberto e, dado o tamanho das principais redes, abordagens que envolvam intervenção manual são impraticáveis. Usuários inativos por longo tempo inviabilizam soluções simples tais como a expiração de um prazo desde o último acesso, o que torna difícil a diferenciação entre inativos e falecidos. Esta pesquisa iniciou-se com o pressuposto de que o problema poderia ser parcialmente resolvido com métodos automáticos e a hipótese era de que dois métodos aqui propostos, um baseado na análise de frequência de mensagens trocadas entre usuários e outro fundamentado na combinação de informações da topologia da rede junto a inspeções de mensagens, poderiam identificar satisfatoriamente parte dos usuários falecidos. Para testar esta hipótese, recorreu-se à simulação computacional, usando topologias livre de escala e aleatória. O programa que simula as redes foi construído de forma a aplicar e testar os métodos de identificação de falecidos, seguindo padrões de projeto que permitem facilmente a troca ou o encadeamento dos algoritmos a validar. Dessa característica, originou-se um terceiro método, que é a combinação das saídas de algoritmos detectores aplicados anteriormente à rede. Os resultados da pesquisa validaram a hipótese, sendo que os dois métodos propostos inicialmente tiveram, cada qual, índices de acerto superiores a 70% na maioria dos casos simulados, independentemente da topologia da rede. Em ambos os métodos, no entanto, é necessária uma calibração de dois parâmetros operacionais, o que exige algum conhecimento da rede examinada e influencia na taxa de detecção. O último método mostrou-se bastante eficiente, com detecção correta superior a 94%, e capaz de absorver flutuações na taxa de detecção dos demais métodos advindas de suas respectivas parametrizações. Portanto, os objetivos da pesquisa foram plenamente atingidos, com a validação da hipótese inicial, a proposta de três métodos para a solução do problema e a geração de um produto tecnológico, o Demortuos, que é o software de simulação da rede e teste dos métodos, atualmente em processo de registro no Instituto Nacional da Propriedade Industrial (INPI). Adicionalmente, foram abertas possibilidades para o desenvolvimento de métodos automáticos para busca de outras classes de usuários<br>Abstract: Identifying deceased users in Online Social Networks is an open challenge and, given the size of the main networks, approaches involving manual intervention are impractical. Inactive users for a long time prevent simple solutions such as the expiration of a period since the last entry, making it difficult to differentiate between inactive and deceased users. This research began with the assumption that the problem could be partially solved with automated methods and the hypothesis was that two methods proposed here, one based on frequency analysis of messages exchanged between users and the other based on the combination of topology information network with the messages of inspections, could satisfactorily identify the part of deceased users. To test this hypothesis, we used the computer simulation, using free topologies of scale and random, the latter for comparison purposes. The program that simulates the network was constructed to implement and test the deceased identification methods, following design patterns that easily allow the exchange or the chain of algorithms to validate. This characteristic gave up a third method, which is combining the outputs of detectors algorithms previously applied to the network. The survey results validated the hypothesis, and the two proposed methods initially had, each, hit rates of over 70% in most cases simulated, regardless of the network topology. In both methods, however, two operating parameters calibration is necessary, which requires some knowledge of the network and examined influences the detection rate. The last method proved to be very efficient with proper detection above 94%, and able to absorb fluctuations in the detection rate of other methods resulting from their respective parameterization. Therefore, the research objectives were fully achieved, with the validation of the initial hypothesis, the proposed three methods for the solution of the problem and the generation of a technological product, Demortuos, which is the network simulation software and testing methods currently in the registration process at the National Institute of Industrial Property (INPI). Moreover, possibilities are opened for the development of automated methods to search for other classes of users<br>Mestrado<br>Tecnologia e Inovação<br>Mestra em Tecnologia

This thesis is devoted to quantitative description, analysis, and modeling of complex social systems in the form of online social networks. Statistical patterns of the systems under study are unveiled and interpreted using concepts and methods of network science, social network analysis, and data mining. A long-term promise of this research is that predicting the behavior of complex techno-social systems will be possible in a way similar to contemporary weather forecasting, using statistical inference and computational modeling based on the advancements in understanding and knowledge of techno-social systems. Although the subject of this study are humans, as opposed to atoms or molecules in statistical physics, the availability of extremely large datasets on human behavior permits the use of tools and techniques of statistical physics. This dissertation deals with large datasets from online social networks, measures statistical patterns of social behavior, and develops quantitative methods, models, and metrics for complex techno-social systems.<br>La presente tesis está dedicada a la descripción, análisis y modelado cuantitativo de sistemas complejos sociales en forma de redes sociales en internet. Mediante el uso de métodos y conceptos provenientes de ciencia de redes, análisis de redes sociales y minería de datos se descubren diferentes patrones estadísticos de los sistemas estudiados. Uno de los objetivos a largo plazo de esta línea de investigación consiste en hacer posible la predicción del comportamiento de sistemas complejos tecnológico-sociales, de un modo similar a la predicción meteorológica, usando inferencia estadística y modelado computacional basado en avances en el conocimiento de los sistemas tecnológico-sociales. A pesar de que el objeto del presente estudio son seres humanos, en lugar de los átomos o moléculas estudiados tradicionalmente en la física estadística, la disponibilidad de grandes bases de datos sobre comportamiento humano hace posible el uso de técnicas y métodos de física estadística. En el presente trabajo se utilizan grandes bases de datos provenientes de redes sociales en internet, se miden patrones estadísticos de comportamiento social, y se desarrollan métodos cuantitativos, modelos y métricas para el estudio de sistemas complejos tecnológico-sociales.

Recentemente, as medidas de centralidade ganharam relevância nas pesquisas com redes complexas e redes sociais, atuando como preditores comportamentais, na identificação de elementos de poder e influência, na detecção de pontos estratégicos para a comunicação e para a transmissão de doenças. Novas métricas foram criadas e outras reformuladas, mas pouco tem sido feito para que se entenda a relação existente entre as diferentes medidas de centralidades, assim como sua relação com outras propriedades estruturais das redes em que elas são frequentemente aplicadas. Nossa pesquisa visa analisar e estudar essas relações para que sirvam de guia na aplicação das medidas de centralidade existentes em novos contextos e aplicações. Nós apresentamos também evidencias que indicam um desempenho superior das medidas conhecidas como Walk Betweenness, Information, Eigenvector and Betweenness na distinção de vértices das redes somente pelas suas características estruturais. Ainda, nós propiciamos detalhes sobre o desempenho distinto de cada métrica de acordo com o tipo de rede em que se trabalha. Adicionalmente, mostramos que várias das medidas de centralidade apresentam um alto nível de redundância e concordância entre si (com correlação superior a 0,8). Um forte indício que o uso simultâneo de várias métricas é improdutivo ou pouco eficaz. Os resultados da nossa pesquisa reforçam a ideia de que para usar apropriadamente as medidades de centralidade é de extrema importância que se saiba mais sobre o comportamento e propriedades das mesmas, fato que salientamos nessa dissertação.<br>Over the last years, centrality measures have gained importance within complex and social networks research, e.g., as predictors of behavior, identification of powerful and influential elements, detection of critical spots in communication networks and in transmission of diseases. New measures have been created and old ones reinvented, but few have been proposed to understand the relation among measures as well as between measures and other structural properties of the networks. Our research analyzes and studies these relations with the objective of providing a guide to the application of existing centrality measures for new environments and new purposes. We shall also present evidence that the measures known as Walk Betweenness, Information, Eigenvector and Betweenness are substantially better than other metrics in distinguishing vertices in a network by their structural properties. Furthermore, we provide evidence that each metric performs better with respect to distinct kinds of networks. In addition, we show that most metrics present a high level of redundancy (over 0.8 correlation) and its simultaneous use, in most cases, is fruitless. The results achieved in our research reinforce the idea that to use centrality measures properly, knowledge about their underlying properties and behavior is valuable, as we show in this dissertation.

In recent years there has been a surge of community detection study of complex network analysis, since communities often play important roles in network systems. Most contemporary community detection algorithms employ single optimization criteria (i.e., modularity), which may not be adequate to represent the structures in complex networks. We suggest a community detection process as a Multi-Objective Optimization Problem (MOP) for investigating the community structures in complex networks. To overcome the limitations of community detection problems, we propose new multi-objective optimization algorithms: a Modified Harmony Search Algorithm, a Hybrid Chaotic Local Search-Harmony Search Algorithm (CLS-HAS) and an Enhanced Firefly Algorithm (EFF). A new tuning parameter based on a chaotic mechanism and novel self-adaptive probabilistic mutation strategies is used to improve the overall performance of the EFF algorithm. Although much of the focus of community detection techniques has been on identifying disjoint and static communities, almost all real networks are dynamic in nature. Detecting communities in dynamic networks is very challenging and the analysis of dynamic communities is still considered to be in its infancy. To study the structure of communities in dynamic networks, we consider an evolution-based clustering method with the aim of maximizing cluster accuracy and minimizing clustering drift from one time step to the next. In this study, the detection of communities with temporal smoothness is formulated as a multi-objective problem and the Modified Bee Swarm Optimization (MBSO) is proposed to solve the community detection problem. The MBSO algorithm uses three kinds of bees, which have different moving pattern, to explore the entire search space and prevent premature convergence. The proposed algorithm has several remarkable characteristics to enhance the search capability of the original bee swarm optimization (BSO) for finding Pareto optimal solutions. Many real networks have complex overlapping community structures. This research also proposes a novel Fungi Optimization Algorithm (FOA) to discover overlapping communities. Unlike conventional algorithms based on node clustering, the proposed algorithm is based on link clustering.