Hibridación del aprendizaje local y el aprendizaje de funciones de distancia
Cargando...
Archivos
Fecha
Autores
Nguyen Cong, Bac
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Central “Marta Abreu” de Las Villas
Resumen
El objetivo general de la investigación consiste en desarrollar un método para construir la función de distancia a partir de restricciones apareadas locales. Luego se adapta el nuevo método a los enfoques locales de múltiples métricas. La incorporación de este método de aprendizaje de distancia es de vital importancia para los investigadores del campo del Aprendizaje Automatizado al contar con nuevos métodos para dar solución a problemas de clasificación. En el contenido del trabajo se expone el marco teórico-referencial de la investigación, enfatizando en las técnicas más empleadas en la actualidad para el aprendizaje de distancia. Se estudia el problema de clasificación basada en instancias usando el paradigma de aprendizaje de la función de distancia de Mahalanobis. Se plantean cuestiones importantes en la escalabilidad y el grado requerido de supervisión de métodos existente de aprendizaje de distancia. Se desarrolla un modelo eficiente de aprendizaje de distancia. También se incorporó una adaptación del algoritmo de los k vecinos más cercanos para dar solución a problemas de clasificación usando distancia de Mahalanobis. Finalmente, se muestra la viabilidad del modelo desarrollado a partir de sus resultados en los conjuntos de datos de Aprendizaje Automatizado reconocido internacionalmente. Se evaluaron, utilizando las pruebas estadísticas no paramétricas. Se demostró de esta forma la hipótesis de investigación planteada.
The general objective of the investigation consists on developing a method to build the distance metric starting from matched up local restrictions. Then the new method adapts to the local focuses of multiple metric. The incorpo-ration of this method of distance metric learning has a high importance for the researches on the field of machine learning because this provides new methods to give solution to classification problems. In the content of the work is exposed theoretical-referential of the investigation, emphasizing in the more used techniques at the present time for the distance metric learning. The classification problem based on instances is studied using the paradigm of the distance metric learning. We plan about important questions in the scalability and the required grade of existent supervision of methods of distance metric learning. An efficient model of distance learning is developed. We also incorporated an adaptation of the algorithm toknearest neighbours to give solution to classification problems using distance of Mahalanobis. Finally, the viability of the approach is shown by its results in the international datasets of Machine Learning. They were evaluated, using the non parametric statistical tests. By this way, the hypothesis of investigation was demonstrated.
The general objective of the investigation consists on developing a method to build the distance metric starting from matched up local restrictions. Then the new method adapts to the local focuses of multiple metric. The incorpo-ration of this method of distance metric learning has a high importance for the researches on the field of machine learning because this provides new methods to give solution to classification problems. In the content of the work is exposed theoretical-referential of the investigation, emphasizing in the more used techniques at the present time for the distance metric learning. The classification problem based on instances is studied using the paradigm of the distance metric learning. We plan about important questions in the scalability and the required grade of existent supervision of methods of distance metric learning. An efficient model of distance learning is developed. We also incorporated an adaptation of the algorithm toknearest neighbours to give solution to classification problems using distance of Mahalanobis. Finally, the viability of the approach is shown by its results in the international datasets of Machine Learning. They were evaluated, using the non parametric statistical tests. By this way, the hypothesis of investigation was demonstrated.