Imagen de Portada

Importance weighted variational graph autoencoder

Guardado en:
Publicado en:Complex & Intelligent Systems vol. 12, no. 1 (Jan 2026), p. 31
Autor principal: Tao, Yuhao
Otros Autores: Guo, Lin, Zhao, Shuchang, Zhang, Shiqing
Publicado:
Springer Nature B.V.
Materias:
Lower bounds
Teaching methods
Learning
Graphical representations
Recommender systems
Graph representations
Optimization
Neural networks
Back propagation
Acceso en línea:Citation/Abstract
Full Text
Full Text - PDF
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
Resumen:Variational Graph Autoencoder (VGAE) is a widely explored model for learning the distribution of graph data. Currently, the approximate posterior distribution in VGAE-based methods is overly restrictive, leading to a significant gap between the variational lower bound and the log-likelihood of graph data. This limitation reduces the expressive power of these VGAE-based models. To address this issue, this paper proposes the Importance Weighted Variational Graph Autoencoder (IWVGAE) and provides a theoretical justification. This method makes the posterior distribution more flexible through Monte Carlo sampling and assigns importance weights to the likelihood gradients during backpropagation. In this way, IWVGAE achieves a more flexible optimization objective, enabling the learning of richer latent representations for graph data. It not only achieves a theoretically tighter variational lower bound but also makes graph density estimation more accurate. Extensive experimental results on seven classic graph datasets show that as the number of samples from the approximate posterior distribution increases, (1) the variational lower bound continuously improves, validating the proposed theory, and (2) the performance on downstream tasks significantly improves, demonstrating more effective learning and representation of graph data.
ISSN:2199-4536
2198-6053
DOI:10.1007/s40747-025-02144-9
Fuente:Advanced Technologies & Aerospace Database