Graph Foundation Models: State of the Art and Future Directions

Abstract

Graph Foundation Models (GFMs) have recently emerged as a promising direction for learning reusable knowledge from graph-structured data. Unlike conventional graph neural networks (GNNs), which are usually trained for a specific graph and task, GFMs aim to transfer graph representations across graphs, tasks, and feature modalities. However, the established literature is rapidly fragmenting in different directions. In this tutorial, we divide current GFM research into five predominant families and explore them appropriately. We first introduce the motivation for GFMs and the high-level taxonomy that organizes the area. We then discuss five representative families: LLM-Enhanced GNNs, GNN-Enhanced LLMs, Unified Graph Models, Graph Mixture-of-Experts, and Graph Prompt Learning illustrating their individual strengths which underpin their current design. Finally, we outline future research directions from a data management perspective. This tutorial will help researchers and practitioners from the database and data management community understand the major implementation paths and open problems in Graph Foundation Models.