Optimizing Graph Neural Network Architecture Through Grammar-Based Genetic Programming

Graph Neural Networks (GNNs) have demonstrated superior capabilities in processing graph-structured data across various domains. However, their deployment is hindered by the time-consuming process of designing effective architectures and optimizing hyperparameters. To address these challenges, this paper introduces a knowledge-driven approach for optimizing high-performance GNNs (KD-GNN). KD-GNN employs grammar-based genetic programming to construct grammar trees guided by domain knowledge and uses a gaussian process-based surrogate model for efficient fitness evaluation. This approach enhances both search efficiency and generalization capabilities, thus making GNN optimization more applicable to a wider range of problems. The paper details the KD-GNN methodology and presents experimental results on graph classification tasks, concluding with insights and future directions.