Use this link to cite:
http://hdl.handle.net/2183/34009 Aprendizaje positivo sin etiquetas (PU Learning) para la mejora de la explicabilidad
Loading...
Identifiers
Publication date
Authors
Fernández-Campa González, Álvaro
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen]: El objetivo de este trabajo consiste en establecer una nueva metodología que mediante
el uso de datos de mayor calidad, permita que los sistemas de recomendación sean más precisos,
explicables y personalizados. En estos sistemas se diferencian muestras positivas, que
son aquellas que ofrecen explicaciones adecuadas que puedan convencer o dar confianza al
usuario en las mismas; y muestras negativas, que son aquellas que no tiene utilidad o no son
adecuadas como explicación de una recomendación. Con el desarrollo de este proyecto se trata
de solucionar la problemática existente en la selección de muestras negativas que se utiliza en
los sistemas del Estado del Arte, donde se asumen como negativas todas aquellas muestras que
no son claramente identificadas como comentarios, imágenes, artículos, etc. pertenecientes al
usuario (positivas). La novedad que se propone es el uso de técnicas de Positive-Unlabeled
(PU) Learning, un modelo de aprendizaje utilizado para situaciones en las que no disponemos
de ejemplos negativos concretos para el entrenamiento. En base a los resultados obtenidos,
se puede afirmar que el empleo de esta técnica a la hora de generar conjuntos de datos negativos
es una manera que nos permite que estos sean más representativos y en definitiva de
mayor calidad. De esta manera, se obtiene un mejor rendimiento de estos sistemas y con ello,
se facilitan mejor explicaciones al entrenar con unos datos con menor ruido en el etiquetado.
[Abstract]: The objective of this work is to establish a new methodology that, through the use of higher quality data, allows recommendation systems to be more accurate, explainable and personalized. In these systems we can differenciate between positive samples, which are those that can convince or provide confidence to the user; and negative samples, which are the ones that do not have utility or are not adequate as an explanation of a recommendation. With the development of this project, it can be addressed the existing issue in the selection of negative samples used in state-of-the-art systems where all samples that are not identified as clearly positive (belonging to the specific user), are assumed to be negative. The novelty proposed is the use of Positive-Unlabeled (PU) Learning techniques, a learning model used in situations where we do not have specific negative examples for training. Based on the results obtained, it can be affirmed that the use of this methodology when generating data sets is a way to make them more representative and ultimately of higher quality. Therefore, both a better performance and better explanations can be obtained with the use of these systems due to the training made with data that contains less noise in the labelling.
[Abstract]: The objective of this work is to establish a new methodology that, through the use of higher quality data, allows recommendation systems to be more accurate, explainable and personalized. In these systems we can differenciate between positive samples, which are those that can convince or provide confidence to the user; and negative samples, which are the ones that do not have utility or are not adequate as an explanation of a recommendation. With the development of this project, it can be addressed the existing issue in the selection of negative samples used in state-of-the-art systems where all samples that are not identified as clearly positive (belonging to the specific user), are assumed to be negative. The novelty proposed is the use of Positive-Unlabeled (PU) Learning techniques, a learning model used in situations where we do not have specific negative examples for training. Based on the results obtained, it can be affirmed that the use of this methodology when generating data sets is a way to make them more representative and ultimately of higher quality. Therefore, both a better performance and better explanations can be obtained with the use of these systems due to the training made with data that contains less noise in the labelling.
Description
Keywords
Inteligencia artificial explicable Explicabilidad basada en imágenes Sistemas de recomendación Aprendizaje automático Filtrado colaborativo ELVis Métodos de PU learning Métricas de calidad de datos Explainable artificial intelligence Image-based recommendation Recommendation systems Machine learning Collaborative filtering PU Learning methods Data quality metrics
Editor version
Rights
Todos los derechos reservados. All rights reserved.







