Use this link to cite:
https://hdl.handle.net/2183/47691 Aplicación de Deep Sets para la representación escalable de usuarios en modelos de explicabilidad visual para sistemas de recomendación
Loading...
Identifiers
Publication date
Authors
García Mateo, Jorge
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen]: En los sistemas recomendadores actuales, la transparencia y la explicabilidad son aspectos clave para aumentar la confianza del usuario. Propuestas como ELVis, MF-ELVis o BRIE han mostrado que es posible proporcionar explicaciones visuales personalizadas a partir de imágenes aportadas por los propios usuarios. Sin embargo, estos enfoques se apoyan en representaciones fijas asociadas a identificadores, lo que introduce una limitación práctica relevante: el sistema debe almacenar un embedding por usuario, su tamaño crece con la base de usuarios y resulta necesario reentrenarlo para integrar nuevos perfiles de forma nativa. Este trabajo propone un camino alternativo orientado a escenarios reales mediante la integración de Deep Sets. El usuario se representa dinámicamente como una agregación de sus imágenes, eliminando la necesidad de mantener una tabla de embeddings. Como consecuencia, el número de parámetros deja de depender del número de usuarios y el tamaño del sistema permanece constante, incluso en plataformas con volúmenes masivos de usuarios. Además, esta formulación permite recomendar a usuarios no vistos durante el entrenamiento siempre que tengan un historial visual previo, sin necesidad de reentrenar el sistema ante la incorporación continua de nuevos perfiles. La propuesta, evaluada como una extensión de BRIE sobre conjuntos de datos reales, muestra un rendimiento competitivo en métricas de ranking y una mejora clara en escalabilidad y capacidad de adaptación. Sin embargo, esta flexibilidad implica un mayor coste computacional en inferencia. En conjunto, los resultados refuerzan que este enfoque es especialmente adecuado para sistemas de recomendación en entornos reales, donde la incorporación continua de nuevos usuarios convierte la escalabilidad en un requisito central del diseño.
[Abstract]: In current recommender systems, transparency and explainability are key aspects for increasing user trust. Approaches such as ELVis, MF-ELVis, or BRIE have shown that it is possible to provide personalized visual explanations from images provided by users themselves. However, these approaches typically rely on fixed representations associated with user identifiers, which introduces a relevant practical limitation: the system must store one embedding per user, its size grows with the user base, and retraining is required to natively integrate new profiles. This work proposes an alternative path oriented towards real-world scenarios through the integration of Deep Sets. Users are represented dynamically as an aggregation of their images, eliminating the need to maintain a user embedding table. As a result, the number of model parameters no longer depends on the number of users, and the system size remains constant even in platforms with massive user volumes. Moreover, this formulation allows recommendations to be provided to users not seen during training, as long as a minimal visual history is available, without requiring retraining when new users are continuously incorporated. The proposed approach, evaluated as an extension of BRIE on real-world datasets, shows competitive performance in ranking metrics together with a clear improvement in scalability and adaptability. However, this flexibility comes at the cost of increased computational effort during inference. Overall, the results reinforce that this approach is particularly suitable for recommender systems in real-world environments, where the continuous arrival of new users makes scalability a central design requirement.
[Abstract]: In current recommender systems, transparency and explainability are key aspects for increasing user trust. Approaches such as ELVis, MF-ELVis, or BRIE have shown that it is possible to provide personalized visual explanations from images provided by users themselves. However, these approaches typically rely on fixed representations associated with user identifiers, which introduces a relevant practical limitation: the system must store one embedding per user, its size grows with the user base, and retraining is required to natively integrate new profiles. This work proposes an alternative path oriented towards real-world scenarios through the integration of Deep Sets. Users are represented dynamically as an aggregation of their images, eliminating the need to maintain a user embedding table. As a result, the number of model parameters no longer depends on the number of users, and the system size remains constant even in platforms with massive user volumes. Moreover, this formulation allows recommendations to be provided to users not seen during training, as long as a minimal visual history is available, without requiring retraining when new users are continuously incorporated. The proposed approach, evaluated as an extension of BRIE on real-world datasets, shows competitive performance in ranking metrics together with a clear improvement in scalability and adaptability. However, this flexibility comes at the cost of increased computational effort during inference. Overall, the results reinforce that this approach is particularly suitable for recommender systems in real-world environments, where the continuous arrival of new users makes scalability a central design requirement.
Description
Keywords
Sistemas de recomendación Explicabilidad visual Conjuntos profundos Representación de usuarios Escalabilidad Aprendizaje automático Orden invariante Eficiencia computacional Recommender systems Visual explainability Deep Sets User representation Scalability Machine learning Permutation invariance Computational efficiency
Editor version
Rights
Attribution-NonCommercial 4.0 International








