Hierarchical classification with bayesian networks and chained classifiers

Hierarchical classification (HC) is a especial type of multi-label classification, that is, there is a set of labels, and instances can be associated to a subset of the labels. Nevertheless, in HC, the labels are arranged in a predefined structure, which is usually tree but in its general form is a...

Descripción completa

Detalles Bibliográficos
Autor: JONATHAN SERRANO PEREZ
Tipo de recurso: tesis de maestría
Estado:Versión aceptada para publicación
Fecha de publicación:2019
País:México
Institución:Instituto Nacional de Astrofísica, Óptica y Electrónica
Repositorio:Repositorio Institucional del INAOE
Idioma:inglés
OAI Identifier:oai:inaoe.repositorioinstitucional.mx:1009/1948
Acceso en línea:http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/1948
Access Level:acceso abierto
Palabra clave:info:eu-repo/classification/Inspec/Hierarchical classification
info:eu-repo/classification/Inspec/Bayesian networks
info:eu-repo/classification/Inspec/Artificial datasets
info:eu-repo/classification/cti/1
info:eu-repo/classification/cti/12
info:eu-repo/classification/cti/1203
info:eu-repo/classification/cti/120323
Descripción
Sumario:Hierarchical classification (HC) is a especial type of multi-label classification, that is, there is a set of labels, and instances can be associated to a subset of the labels. Nevertheless, in HC, the labels are arranged in a predefined structure, which is usually tree but in its general form is a Directed Acyclic Graph (DAG). Furthermore, in HC there are different problems, which can be described by the type of hierarchical structure, the number of paths which an instance can be associated and the depth of the paths. In this work is proposed a method for hierarchical classification, which can handle tree and DAG hierarchies, and predicts a single path which always reaches a leaf node. The method takes advantage of the hierarchical structure to inuence the prediction of local classifiers with their neighbors, to achieve this, two different strategies are combined. The first is to represent the hierarchical structure as a Bayesian network, which represents the data distribution in the nodes while maintains the hierarchical constraint; the second is to train chained classifiers, that feed the Bayesian network, in this way, the classifiers are considering the hierarchical structure. Furthermore, four different variants of the method were implemented, the main difference between them is the neighbors that inuence the predictions on the chained classifiers. Due to the different hierarchical classification problems, the real world datasets for each problem are limited. So, a way to evaluate or extend the analysis of a method is to generate Artificial Datasets (AD). Thus, a method to build artificial datasets for different hierarchical classification problems is proposed. The method generates instances from the distribution of each node, so, it requires as input the distribution for each leaf node, the distributions for internal nodes are estimated by the method. In this way, several artificial datasets have been generated, which are divided in two main groups, those with hierarchy tree type and those with hierarchy DAG type. Both groups were made available to the scientific community. Finally, the different variants of the proposed method for HC were evaluated with real world and artificial datasets. Later, their results were compared against standard and state of the art methods, then all the results were analyzed with Friedman test and its post-hoc the Nemenyi test.