Go to ICLR 2025 Conference homepageGraphProp: Training the Graph Foundation Models using Graph Properties
In this work, we focus on training Graph Foundation Models (GFMs) for graph-level tasks like protein classification. Effective GFM training requires capturing information consistent across different domains. We have discovered that graph structures provide more consistent cross-domain information compared to node features and graph labels. However, traditional in-context learning methods primarily focus on transferring node features from various domains into a unified representation space but often lack structural cross-domain generalization. To address this, we introduce a method called GraphProp, which emphasizes structural generalization. The GraphProp training process consists of two main phases: initially, it trains a structural GFM through the supervised prediction of graph structural properties. It then uses the structural representation from this GFM as positional encoding to train a comprehensive GFM. This phase of training utilizes in-context learning with domain-specific node features and graph labels to improve cross-domain node feature generalization. Additionally, employing data augmentation in training the structural GFM helps address the scarcity of labeled graph data and facilitates explicit cross-domain structural generalization. Our experimental results demonstrate that GraphProp significantly outperforms traditional in-context learning methods, especially in handling graphs without node features.