Self-organization map in complex network

• Autor(a) principal: Mayra Mercedes Zegarra Pimenta
• Data de Publicação: 2018
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://doi.org/10.11606/T.55.2018.tde-30102018-111955

The Self-Organization Map (SOM) is an artificial neural network that was proposed as a tool for exploratory analysis in large dimensionality data sets, being used efficiently for data mining. One of the main topics of research in this area is related to data clustering applications. Several algorithms have been developed to perform clustering in data sets. However, the accuracy of these algorithms is data depending. This thesis is mainly dedicated to the investigation of the SOM from two different approaches: (i) data mining and (ii) complex networks. From the data mining point of view, we analyzed how the performance of the algorithm is related to the distribution of data properties. It was verified the accuracy of the algorithm based on the configuration of the parameters. Likewise, this thesis shows a comparative analysis between the SOM network and other clustering methods. The results revealed that in random configuration of parameters the SOM algorithm tends to improve its acuracy when the number of classes is small. It was also observed that when considering the default configurations of the adopted methods, the spectral approach usually outperformed the other clustering algorithms. Regarding the complex networks approach, we observed that the network structure has a fundamental influence of the algorithm accuracy. We evaluated the cases at short and middle learning time scales and three different datasets. Furthermore, we show how different topologies also affect the self-organization of the topographic map of SOM network. The self-organization of the network was studied through the partitioning of the map in groups or communities. It was used four topological measures to quantify the structure of the groups such as: modularity, number of elements per group, number of groups per map, size of the largest group in three network models. In small-world (SW) networks, the groups become denser as time increases. An opposite behavior is found in the assortative networks. Finally, we verified that if some perturbation is included in the system, like a rewiring in a SW network and the deactivation model, the system cannot be organized again. Our results enable a better understanding of SOM in terms of parameters and network structure.