Seleção de componentes em ensembles de clasificadores multirrótulo / Component Selection in Ensembles of Multi-label Classifiers
AUTOR(ES)
NATHANAEL DE CASTRO COSTA
FONTE
IBICT - Instituto Brasileiro de Informação em Ciência e Tecnologia
DATA DE PUBLICAÇÃO
27/07/2012
RESUMO
The selection of components in ensembles of classifiers is a very common activity in the field of Machine Learning with several studies showing its effectiveness in obtaining significant gains in accuracy. However, the most studied classification task involves mutually exclusive labels (classes). The objective of this work is to present a study on the selection of components in ensembles of multi-label classifiers, whereby an instance can become associated with more than one label. Two search approaches for the component selection were used, one based on Genetic Algorithms and the other based on Hill Climbing. Conversely, two types of evaluation measures were adopted for ensemble selection: one based on multilabel accuracy measures and the other based on well known diversity measures for ensembles, which were modified to handle the multilabel case. Another selection approach was also conceived to assign different ensembles to different labels of the labelset. Specifically for generating the RAkEL components, 10 well known learning algorithms for inducing one-label classifiers were used, namely, Naïve Bayes, RBF Neural Networks, Support Vector Machines, J48, REP Tree, IBk, Decision Stump, OneR, PART, and Decision Table. A systematic empirical study was conducted on seven publicly available datasets, involving two ensemble models, each making use of one of the well known multi-label classifiers RAkEL and ML-RBF. In general, the results achieved show that the ensembles produced via ensemble selection can yield better results than the full ensembles and some of their components. The ensemble selection type based on multilabel accuracy measures performed usually better than the other based on diversity measures. Finally, among the search procedures for conducting the selection, none has prevailed over the other. Keywords: Machine Learning, Multi-label Classification, Ensembles, Ensemble Selection, Diversity Measures, Genetic Algorithms, Hill Climbing.
ASSUNTO(S)
algorÍtmos genÉticos - dissertaÇÕes aprendizado computacional - dissertaÇÕes sistemas de informacao redes neurais - dissertaÇÕes
ACESSO AO ARTIGO
http://www.unifor.br/tede//tde_busca/arquivo.php?codArquivo=897721Documentos Relacionados
- Learning classifier system for multi-label classification
- Abordagens para aprendizado semissupervisionado multirrótulo e hierárquico
- Classificação automática de páginas Web Multi-label via MDL e Support Vector Machines
- Hierarchical multilabel classification techniques
- Geração, seleção e combinação de componentes para ensembles de redes neurais aplicadas a problemas de classificação