EGNAS: Efficient Graph Neural Architecture Search through Evolutionary Algorithm

Abstract

The integration of Neural Architecture Search (NAS) techniques into Graph Neural Networks (GNNs) has demonstrated promising outcomes. In this paper, we present the Efficient Graph Neural Architecture Search (EGNAS) algorithm, which aims to tackle the challenges in GNN NAS. EGNAS introduces novel techniques, including inherited parameter sharing and the half epochs technique, to improve the efficiency and diversity of architecture search. Additionally, combined evolutionary search with hyperparameter is incorporated to explore a wider range of GNN candidates. Experimental results demonstrate that EGNAS outperforms handcrafted methods and other GNN NAS approaches in terms of both test accuracy and search time. By striking a balance between performance and computational efficiency, EGNAS offers an efficient and effective solution for discovering optimal graph neural architectures.