Improved search in graph AutoML: expansion and dynamic prioritization in the search space for enhanced efficiency

This paper explores methods for enhancing the automated architecture search process for graph neural networks. We propose a novel approach that dynamically selects a priority direction within the search space, improving the efficiency and quality of the discovered architectures. Another proposed approach expands the search space by allowing combinations of different types of graph convolutional layers. The primary focus is on maximizing the quality of architectures within the expanded search space while maintaining a fixed search budget in terms of the number of models. Our experiments are conducted on datasets from citation networks, chemical molecules, and shopping graph domains. The experimental results show that the proposed approach enables the discovery of more effective and higher-quality models without increasing computational resources, demonstrating high potential for automating solutions to real-world graph data analysis tasks.