Por favor, use este identificador para citar o enlazar este ítem: http://inaoe.repositorioinstitucional.mx/jspui/handle/1009/2691
Semi-Supervised Hierarchical Classification
Jonathan Serrano Pérez
Luis Enrique Sucar
Acceso Abierto
Atribución-NoComercial-SinDerivadas
Hierarchical classification
Semi-supervised learning
Bayesian networks
Transfer learning
Hierarchical classification is commonly seen as a special type of multi-label classification, where the instances can be associated to multiple labels, but labels are arranged in a predefined structure, a hierarchy. The predictions in hierarchical classification have to fulfill the hierarchical constraint that states if an instance is associated to a node, then it also has to be associated to the ancestors of that node. Moreover, scarcity of labeled data is a common problem in supervised classification, because hand-labeling data is expensive, time-consuming or difficult to label; it is a problem also present in hierarchical classification. Even though, labels arranged in hierarchies are found in multiples domains, such as text categorization, image classification, biology and music, just a few works address the problem of scarcity of labeled data in a hierarchical classification scenario. Therefore, the main goal of this research it to develop a semi-supervised hierarchical classifier that can be trained with labeled and unlabeled data, for the hierarchical problem where the hierarchies can be of directed acyclic graph type and the instances can be associated to multiple paths of labels of partial depth. Semi-supervised hierarchical multi-label classifier based on local information (SSHMC-BLI) is the proposed classifier, which can be trained with labeled and unlabeled data to perform hierarchical classification tasks. The method mainly consists on building pseudo-labels for each unlabeled instance from the paths of labels of its labeled neighbors, while it considers whether the unlabeled instance is similar to its neighbors. Experiments on several artificial and real world datasets show that using unlabeled along with labeled data can help to improve the performance of a supervised hierarchical classifier trained only on labeled data, with statistical significance. Finally, some extensions of SSHMC-BLI were proposed: SSHC-BLI which is variant that handles only hierarchies of tree type; SSHBMC which follows a Bayesian approach, that is, the hierarchy is modeled as a Bayesian network; and a variant of SSHMC-BLI that incorporates transfer learning among neighboring nodes.
Instituto Nacional de Astrofísica, Óptica y Electrónica
2025-04
Tesis de doctorado
Inglés
Estudiantes
Investigadores
Público en general
Serrano Pérez, J., (2025), Semi-Supervised Hierarchical Classification, Tesis de Doctorado, Instituto Nacional de Astrofísica, Óptica y Electrónica
OTRAS ESPECIALIDADES TECNOLÓGICAS
Versión aceptada
acceptedVersion - Versión aceptada
Aparece en las colecciones: Maestría en Ciencias Computacionales

Cargar archivos:


Fichero Tamaño Formato  
SERRANOPJ_DCC.pdf7.36 MBAdobe PDFVisualizar/Abrir