Seleção e rotulagem de instâncias para métodos semissupervisionados indutivos
In recent years, the use of Machine Learning (ML) techniques to solve real problems has become very common and a technological pattern adopted in plenty of domains. However, several of these domains do not have enough labelled data to give ML methods a good performance. This problem led to the de...
Na minha lista:
Autor principal: | |
---|---|
Outros Autores: | |
Formato: | doctoralThesis |
Idioma: | pt_BR |
Publicado em: |
Universidade Federal do Rio Grande do Norte
|
Assuntos: | |
Endereço do item: | https://repositorio.ufrn.br/handle/123456789/55155 |
Tags: |
Adicionar Tag
Sem tags, seja o primeiro a adicionar uma tag!
|
Resumo: | In recent years, the use of Machine Learning (ML) techniques to solve real problems has
become very common and a technological pattern adopted in plenty of domains. However,
several of these domains do not have enough labelled data to give ML methods a good
performance. This problem led to the development of Semi-supervised methods, a type of
method capable of using labelled and unlabelled instances in its model building. Among
the semi-supervised learning techniques, the inductive methods stand out. The wrapper
methods, a particular category within inductive methods, use a process, often iterative,
that involves: training the method with labelled data; selection of the best data from the
unlabelled set; and labelling the selected data. Despite showing a simple and efficient
process, errors in the selection or labelling processes are common, which deteriorate the
final performance of the method. This research aims to reduce selection and labelling errors
in wrapper methods to establish selection and labelling approaches that are more robust
and less susceptible to errors. To this end, this work proposes a selection and labelling
approach based on classification agreement and a selection and agreement approach based
on distance metric as an additional factor to an already used selection criterion (e.g.
confidence or agreement). The proposed approaches can be applied to any wrapper method
and were tested on 42 datasets with Self-training, Co-training and Boosting methods. The
results obtained indicate that the proposals bring gains for both methods in terms of
accuracy and F-measure. |
---|