Algoritmos dinámicos para el agrupamiento con traslape

Clustering is a Data Mining and Machine Learning technique that has been used in several areas like medicine, marketing, social network analysis and bioinformatics, among others. Although, several clustering algorithms have been proposed, they have some drawbacks that are solved through this PhD res...

Descripción completa

Detalles Bibliográficos
Autor: AIREL PEREZ SUAREZ
Tipo de recurso: tesis doctoral
Estado:Versión aceptada para publicación
Fecha de publicación:2011
País:México
Institución:Instituto Nacional de Astrofísica, Óptica y Electrónica
Repositorio:Repositorio Institucional del INAOE
Idioma:español
OAI Identifier:oai:inaoe.repositorioinstitucional.mx:1009/714
Acceso en línea:http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/714
Access Level:acceso abierto
Palabra clave:info:eu-repo/classification/Minería de datos/Data mining
info:eu-repo/classification/Teoría del algoritmo/Algorithm theory
info:eu-repo/classification/Teoría de grafos/Graph theory
info:eu-repo/classification/cti/1
info:eu-repo/classification/cti/12
info:eu-repo/classification/cti/1203
Descripción
Sumario:Clustering is a Data Mining and Machine Learning technique that has been used in several areas like medicine, marketing, social network analysis and bioinformatics, among others. Although, several clustering algorithms have been proposed, they have some drawbacks that are solved through this PhD research. Most clustering algorithms do not allow building overlapping clusterings. However, there are several applications like topics detection, document segmentation, information organization and news analysis, among others, where it is common for objects to belong to more than one cluster; these applications need clustering algorithms able to build overlapping clusters. The majority of the current overlapping clustering algorithms have some drawbacks which can reduce their usefulness in practical applications. These limitations are mainly related with: (a) the necessity of tuning several parameters whose values depend on the collection to cluster, (b) the production of a large number of clusters, usually with a low average of elements per cluster, and (c) the production of clusters with high overlapping. Besides, most of overlapping clustering algorithms are unable to satisfy new requirements such as: (i) the necessity of updating the clustering when the collection changes and (ii) the necessity of building hierarchies of clusters, in which the overlapping among the clusters of the same level are allowed. This PhD research introduces two new overlapping clustering algorithms, DClustR and DHClustR, both satisfy the requirements above mentioned and solve the limitations a), b) and c). DClustR is a dynamic and non-hierarchical algorithm, based on concepts of graph theory, which builds a set of overlapping clusters. DClustR introduces a new strategy for building the clustering and also it introduces a new strategy for updating this clustering when the collection changes, due to multiple additions, eliminations or modifications of objects. On the other hand, DHClustR is a dynamic agglomerative hierarchical clustering algorithm which builds a hierarchy of overlapping clusters, using the DClustR algorithm for building the clustering of each level. For building the clustering of the first level, DClustR is applied over the collection of objects; from this point on, the objects to be clustered at each level are the clusters of the previous level.