@PhDThesis{Parreira:2017:MéClHi,
author = "Parreira, Michelle de Oliveira",
title = "HSMI: m{\'e}todo de classifica{\c{c}}{\~a}o hier{\'a}rquico
baseado em SVM multikernel com otimiza{\c{c}}{\~a}o
meta-heur{\'{\i}}stica",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2017",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2016-11-24",
keywords = "m{\'a}quinas de vetores de suporte, combina{\c{c}}{\~a}o de
classificadores, classifica{\c{c}}{\~a}o bin{\'a}ria,
sensoriamento remoto, reconhecimento de padr{\~o}es, support
vector machine, ensembles, binary classification, remote sensing,
pattern recognition.",
abstract = "Esse trabalho prop{\~o}e o m{\'e}todo HSMI (Hierarchical Support
vector machine with Multiple kernels optimized by Invasive weed
optimization) de classifica{\c{c}}{\~a}o baseado em
m{\'a}quinas de vetores suporte (SVM) que usa m{\'u}ltiplos
kernels e atribui os r{\'o}tulos {\`a}s classes de modo
hier{\'a}rquico. Uma {\'a}rvore bin{\'a}ria {\'e} criada
automaticamente pelo algoritmo proposto e cada n{\'o} realiza a
classifica{\c{c}}{\~a}o entre duas parti{\c{c}}{\~o}es do
conjunto de classes pr{\'e}-classificado pelo n{\'o} superior. A
classifica{\c{c}}{\~a}o {\'e} realizada pelo classificador SVM
com m{\'u}ltiplos kernels combinados aproveitando as diferentes
caracter{\'{\i}}sticas de cada kernel. A escolha pelas classes
que comp{\~o}em cada parti{\c{c}}{\~a}o em cada n{\'o} {\'e}
feita por otimiza{\c{c}}{\~a}o junto com os par{\^a}metros dos
kernels e os coeficientes da combina{\c{c}}{\~a}o linear entre
eles. Para isso {\'e} empregado o algoritmo
Infesta{\c{c}}{\~a}o por Ervas Daninhas (Invasive Weed
Optimization, IWO). Esse novo m{\'e}todo consegue separar
hierarquicamente as classes com melhor separabilidade segundo um
classificador SVM multikernel otimizado para cada
classifica{\c{c}}{\~a}o bin{\'a}ria. Os resultados foram
comparados com o m{\'e}todo SVM com kernel gaussiano e SVM com
kernel polinomial. Os resultados demonstraram que o m{\'e}todo
HSMI ao particionar as classes de forma embutida permite a
fus{\~a}o de classes confusas identificadas no processo de
classifica{\c{c}}{\~a}o. ABSTRACT: This work proposes the
classifier method HSMI (Hierarchical support vector machine with
multiple kernels optimized by Invasive weed optimization) based on
support vector machine (SVM) that uses multiple kernels and
assigns the labels to classes in a hierarchical way. A binary tree
is automatically created by the proposed algorithm and each node
performs the classification between two partitions of the set of
pre-sorted classes by the upper node. The classification is
performed by the SVM classifier with multiple kernels combined
taking advantage of the different characteristics of each kernel.
The choice of the classes that make up each partition at each node
is done by optimization along with the parameters of the kernels
and the coefficients of the linear combination between them. For
this the Invasive Weed Optimization algorithm (IWO) is used. This
new method can separate hierarchically classes with better
separability according to a multi-kernel SVM classifier optimized
for each binary classification. The results were compared with the
SVM method with Gaussian kernel and SVM with polynomial kernel.
The results showed that the HSMI method in partitioning the
classes of embedded form allows the fusion of confused classes
identified in the classification process.",
committee = "Santos, Rafael Duarte Coelho dos (presidente) and Dutra, Luciano
Vieira (orientador) and Pantale{\~a}o, Eliana (orientadora) and
Forster, Carlos Henrique Quartucci and Negri, Rog{\'e}rio Galante
and Mascarenhas, Nelson Delfino d'{\'A}vila",
copyholder = "SID/SCD",
englishtitle = "HSMI: hierarchical classification method based on multi-kernel SVM
with meta-heuristic optimization",
language = "pt",
pages = "111",
ibi = "8JMKD3MGP3W34P/3N7M8QP",
url = "http://urlib.net/ibi/8JMKD3MGP3W34P/3N7M8QP",
targetfile = "publicacao.pdf",
urlaccessdate = "28 abr. 2024"
}