Automated graph neural network architecture selection via a hypothesis-testing loop with large language models

Today, when applying graph neural networks, the problem of selecting an appropriate architecture is relevant. Existing approaches to architecture search and hyperparameter tuning are typically resource-intensive and opaque in terms of the decisions they make, while emerging methods based on large language models offer a promising way to address this gap. The goal is to develop and experimentally evaluate an algorithm for automatic graph neural network architecture selection based on a "hypothesis–measurement–decision–reflection" reasoning loop, in which hypotheses are generated by a large language model and verified in a controlled Python environment with no restrictions on the set of analyzed graph properties. The proposed HypoGNN algorithm formalizes architecture search as an iterative decision-making process. Each iteration consists of four steps with mandatory fixation of the formulated hypothesis, measurement of graph properties, training of the architecture, and reflection on the step's outcome. A systematic comparison of HypoGNN with six baseline methods under a comparable computational budget is carried out. The proposed method achieves the best average accuracy on five out of six datasets and demonstrates the most stable performance among all considered approaches. The formal structure of the loop also enables quantitative analysis of the search trajectory. The obtained results can be used in the development of applied models on graph data. The proposed algorithm can be integrated into existing graph neural network training pipelines.