Graph Foundation Models Re-evaluated for Node Property Prediction

Graph-structured data is prevalent across various industries and scientific fields, leading to significant interest in the development of Graph Foundation Models (GFMs). A particularly popular application for GFMs is node property prediction, which has wide-ranging real-world uses from fraud detection in financial networks to recommendation systems in e-commerce. Despite the proliferation of GFMs for this task, the field lacks a standardized evaluation framework, making reliable comparisons between different GFM approaches and traditional Graph Neural Network (GNN) baselines challenging. This research addresses the issue by conducting a fair and rigorous reevaluation of nine recently proposed GFMs specifically designed for node property prediction. These GFMs were compared against strong, well-tuned GNN baselines. The findings indicate that, among the GFMs assessed, only the very latest models—those built upon the Prior-data Fitted Networks paradigm—demonstrate superior predictive performance compared to optimized GNNs. However, this improved performance often comes at the cost of higher inference complexity. The study underscores the critical need for unified evaluation settings to foster more accurate and dependable comparisons within the rapidly evolving field of Graph Machine Learning.