An Approach Using BRKGA for Optimizing Graph Neural Network Architectures
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Graph Neural Networks (GNNs) have emerged as powerful tools for modeling graph-structured data, demonstrating success in node classification, link prediction, and graph classification. However, designing optimal GNN architectures remains a challenging and resource-intensive task due to the complexity of hyperparameter choices and the high-dimensional nature of graph data. Neural Architecture Search (NAS) has been proposed as a solution to automate this process, with evolutionary algorithms showing particular promise in exploring large architectural spaces. This paper introduces the Biased Random-Key Genetic Algorithm for Graph Neural Networks (BRKGA-GNN), a novel framework that applies BRKGA to optimize GNN architectures. BRKGA offers a robust and efficient search mechanism for navigating the architecture space by leveraging biased random keys and evolutionary strategies. To validate the effectiveness of BRKGA-GNN, we conducted experiments on three widely used benchmark datasets, comparing their performance against state-of-the-art NAS-based methods. The proposed methodology provides an automated and scalable solution for designing effective GNN architectures while minimizing reliance on manual tuning. This study highlights the potential of BRKGA-GNN to advance research in both NAS and GNN optimization, inviting further exploration of its applications and benefits.