Enterprise AI Analysis

Proficient Graph Neural Network Design by Accumulating Knowledge on Large Language Models

This paper introduces DesiGNN, a knowledge-centered framework for automating the design of data-aware Graph Neural Networks (GNNs) for unseen tasks. It addresses limitations in current LLM-based GNN design, such as external noise from descriptive inputs and inherent knowledge gaps in fine-grained design. DesiGNN systematically converts past model design experiences into structured, fine-grained knowledge priors, well-suited for meta-learning with LLMs. By aligning empirical property filtering from extensive benchmarks with adaptive elicitation of literature insights via LLMs, it builds a solid meta-knowledge between graph topology and high-performing GNN architectures. This enables DesiGNN to deliver immediate, data-aware GNN recommendations without any training on unseen data, rapidly refining these architectures based on accumulated insights, and ultimately delivering superior performance with minimal computational overhead. Experiments show DesiGNN delivers initial GNN proposals ranking in the top 5.77% of all possible architectures within seconds and achieves consistently superior performance with minimal search cost compared to baselines.