Two-level domain-decomposition AdaGrad method for scalable training of graph neural networks

DOI10.48550/arXiv.2608.22575MaRDI QIDQ7285075arXivFDO

Authors Julien Herrmann, Alexander Heinlein, Serge Gratton, Alena Kopaničáková, Laurynas Varnas

Publication date 23 August 2026

Published in arXiv

Tldr We propose two novel domain-decomposition variants (DD-AG2m and 2DD-AG2m) of an AdaGrad method enhanced with second-order curvature information and momentum, for scalable training of graph neural networks in distributed environments; these methods reduce computational cost by a factor of 4-8 while improving predictive performance by up to 22% compared to the baseline.

Full work available at URL https://doi.org/10.48550/arXiv.2608.22575

arXiv classification

cs.LGmath.NA

Cites work

• An introduction to domain decomposition methods. Algorithms, theory, and parallel implementation
• Complexity and performance for two classes of noise-tolerant first-order algorithms
• Data-parallel neural network training via nonlinearly preconditioned trust-region method
• Enhancing training of physics-informed neural networks using domain decomposition-based preconditioning strategies
• Globally Convergent Multilevel Training of Deep Residual Networks
• Iterative Methods for Sparse Linear Systems
• Machine learning and domain decomposition methods -- a survey
• Multi-Preconditioned LBFGS for Training Finite-Basis PINNs
• Multilevel Objective-Function-Free Optimization with an Application to Neural Networks Training
• PyTorch: An Imperative Style, High-Performance Deep Learning Library
• The Graph Neural Network Model
• Trust Region Methods
• Trust-region algorithms for training responses: machine learning methods using indefinite Hessian approximations

This page was built for publication: Two-level domain-decomposition AdaGrad method for scalable training of graph neural networks

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7285075)