GraphAgent: An Effective Knowledge-Guided GNN Model Selection System

Abstract

We demonstrate GraphAgent, a user-friendly and effective system for knowledge-guided selection of Graph Neural Network (GNN) models for new graph data. Existing methods, such as Neural Architecture Search, are computationally expensive, while pretrained graph foundation models often fail to generalize across diverse datasets and tasks. GraphAgent addresses these challenges by unifying over 413,000 performance records from public GNN benchmarks into a comprehensive Graph-Model Knowledge Graph (GMKG), encoding relationships between datasets, models, and tasks via ranking-based and performance-weighted edges. A reinforcement learning-based feature selector identifies the most informative metafeatures for each dataset, and model features are derived from normalized performance patterns. A Performance-Aware heterogeneous GNN, trained with listwise ranking loss, propagates structural and performance signals across the GMKG, enabling accurate model ranking prediction for new data. Given a user-uploaded graph and task, GraphAgent rapidly selects the top- k models with interpretable reasoning. Extensive evaluations show GraphAgent achieves up to 95.4% nDCG@1 efficiently.