A Two-Stage Particle Competition Model for Unbalanced Community Detection in Complex Networks

• Ano de defesa: 2020
• Autor(a) principal: Martins, Luan Vinicius de Carvalho
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Community detection; Competição de partícula; Complex networks; Comunidades desbalanceadas; Detecção de comunidade; Particle competition; Rede complexa; Unbalanced community
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-20082020-101929/

The usage of Complex Networks has proved to be an excellent tool in reveling prevalent information from complex systems due to its ability to describe spatial, functional, and topological relations among the data. One inherent characteristic of Complex Networks, which is an excellent source of information, is its community structurecommonly defined as a set of nodes more densely connected than to other nodes of the networks. In order to extract this information, many techniques have been proposed. One interesting technique is the Particle Competition method, which is a bio-inspired approach in which a set of particles are inserted into the network and must compete with themselves to capture as many nodes as possible. Competition, here represented as a stochastic dynamic system that controls the particles, is a behavior widely encountered in nature when there is a shortage of resources, such as water, food, or matesthe nodes of the graph are the limited resources each particle must conquer. However, unbalanced communities commonly appear in real complex networks. Although many community detection techniques have been developed, and some of them possess a certain degree of tolerance to unbalance, there is still lacking an explicit and efficient mechanism to treat this problem. In this document, we proposed a Two-Stage Particle Competition model to detect unbalanced communities. At the first stage, named Competition, the particles compete with themselves to occupy as many nodes as possible. At the second stage, a diffusion-like regularization mechanism is introduced to determine the dominance level of each particle at a neighborhood of each node. The two stages work alternatively until the regularization process converges. In the original Particle Competition model, all particles have the same behavior; therefore, there is no way to correctly occupy the communities with different sizes or structures by the particles. In the proposed model, the regularization mechanism makes each particle to have a different behavior according to the network structure. Consequently, communities with different sizes or structures can be correctly detected by the particles. Computer simulations show promising results of the proposed model. Moreover, the regularization mechanism improves both the accuracy and computational speed of the method as fewer iterations are required until convergence when compared to previous Particle Competition methods