Título del libro: Proceedings Of The International Joint Conference On Neural Networks
Título del capítulo: Cyclo-VGAE: Dual-Mechanism Approach to GNN Robustness Against Noisy Labels
CARLOS MINUTTI MARTINEZ; BORIS ESCALANTE RAMIREZ; JIMENA OLVERES MONTIEL;
Autores externos:
Idioma:
Año de publicación:
2025
Palabras clave:
Adversarial Augmentation; Ensembles; Graph Neural Networks; Label Noise; ogbg-ppa; Protein-Protein Association Network; Robustness; Variational Graph Auto-Encoders
Resumen:
© 2025 IEEE.Graph Neural Networks (GNNs) have become indispensable in learning from graph-structured data, with applications ranging from social network analysis to bioinformatics. However, their performance is significantly degraded by noisy labels, which are common in real-world datasets. This paper addresses the challenge of label noise in GNNs through two approaches: (1) an adaptation of ExpC+bot, one of the top-performing models on the clean data that we optimize extensively for noise robustness, and (2) our novel Cyclo-VGAE method, which we introduce in this work and which won the 2025 International Joint Conference on Neural Networks (IJCNN) Competition:'Learning with Noisy Graph Labels'. Cyclo-VGAE employs Variational Graph Auto-Encoders (VGAE) with a statistical ensemble approach specifically designed to mitigate label noise. For the ExpC+bot adaptation, we conduct comprehensive hyperparameter optimization and component modifications to enhance its performance in noisy settings, while preserving its original architecture. We compare both approaches on the Protein-Protein Association Network dataset (ogbg-ppa) with varying levels of label noise. Our results demonstrate that our proposed Cyclo-VGAE outperforms the adapted ExpC+bot, achieving higher F1 scores and demonstrating greater stability across different hyperparameter configurations. This work provides valuable insight into the handling of noisy labels in GNNs and offers a novel solution to this challenging problem.
Entidades citadas de la UNAM:
ISBN:
9798331510428
Editorial:
Institute of Electrical and Electronics Engineers Inc.
