Use this link to cite:
http://hdl.handle.net/2183/41317 Personalización, explicabilidad y sostenibilidad en sistemas de recomendación
Loading...
Identifiers
Publication date
Authors
Esteban Martínez, David
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen]: Este trabajo aborda la problemática de la transparencia en los Sistemas de Recomendación (SR), proponiendo técnicas innovadoras para mejorar la explicabilidad visual, con un enfoque en la sostenibilidad. Para ello, hemos llevado a cabo un estudio con tres modelos del Estado del Arte (ELVis, MF-ELVis y BRIE) aplicados a reseñas de Tripadvisor. Las técnicas empleadas se centran en optimizar la calidad de los datos de entrada, a través de tres estrategias principales: 1) Selección de nuevos positivos y negativos con Aprendizaje Positivo y Sin Etiquetas, 2) Aumentación de datos por transformación de imágenes o IA Generativa, y 3) Mejora de embeddings de imágenes. Los resultados demuestran que la mejora de embeddings incrementa el rendimiento en un 30%, reduce el tiempo de entrenamiento, y disminuye las emisiones y el consumo en un 20% y 15% en entrenamiento e inferencia, respectivamente. La combinación con las demás técnicas genera una mejora adicional del 5% en el rendimiento y reduce significativamente las épocas de entrenamiento, disminuyendo las emisiones hasta un 60% adicional. Finalmente, estas técnicas han sido incluidas en una librería pública de carácter modular, permitiendo su utilización en entornos de producción, así como en investigaciones incrementales que partan de este trabajo.
[Abstract]: This work addresses the issue of transparency in Recommender Systems (RS), proposing innovative techniques to improve visual explainability, with a focus on sustainability. To this end, we conducted a study with three state-of-the-art models (ELVis, MF-ELVis, and BRIE) applied to Tripadvisor reviews. The employed techniques focus on optimizing the quality of input data through three main strategies: 1) Selection of new positives and negatives using Positive and Unlabeled Learning, 2) Data augmentation through image transformation or Generative AI, and 3) Enhancement of image embeddings. The results show that improving embeddings increases performance by 30%, reduces training time, and decreases emissions and consumption by 20% and 15% in training and inference, respectively. The combination with the other techniques provides an additional 5% performance improvement and significantly reduces training epochs, lowering emissions by up to an additional 60%. Finally, these techniques have been included in a publicly available modular library, allowing their use in production environments as well as in incremental research building upon this work.
[Abstract]: This work addresses the issue of transparency in Recommender Systems (RS), proposing innovative techniques to improve visual explainability, with a focus on sustainability. To this end, we conducted a study with three state-of-the-art models (ELVis, MF-ELVis, and BRIE) applied to Tripadvisor reviews. The employed techniques focus on optimizing the quality of input data through three main strategies: 1) Selection of new positives and negatives using Positive and Unlabeled Learning, 2) Data augmentation through image transformation or Generative AI, and 3) Enhancement of image embeddings. The results show that improving embeddings increases performance by 30%, reduces training time, and decreases emissions and consumption by 20% and 15% in training and inference, respectively. The combination with the other techniques provides an additional 5% performance improvement and significantly reduces training epochs, lowering emissions by up to an additional 60%. Finally, these techniques have been included in a publicly available modular library, allowing their use in production environments as well as in incremental research building upon this work.
Description
Keywords
Aprendizaje automático Sostenibilidad Inteligencia artificial explicable Inteligencia artificial verde Inteligencia artificial frugal Aprendizaje positivo y sin etiquetas Aumentación de los datos Sistemas de recomendación Inteligencia artificial generativa Machine learning Sustainability Explainable artificial intelligence Green artificial intelligence Frugal artificial intelligence Positive unlabeled learning Data augmentation Recommender systems Generative artificial intelligence
Editor version
Rights
Atribución 3.0 España







