Work

Text-Free Multi-domain Graph Pre-training: Toward Graph Foundation Models

Year: 2026
Type: article
Abstract: Graphs arise across a wide spectrum of real-world applications, spanning biology, e-commerce, social networks, and more. This ubiquity naturally raises a fundamental question: Can we build a graph foundation model that leverages knowledge from multiple domains and generalizes effectively to entirely new ones? Achieving this goal, however, is challenging due to the substantial variation in graph characteristics across domains. While early efforts have explored cross-domain graph pre-training, many rely on auxiliary textual data to bridge semantic gaps between graphs, limiting their applicability to text-attributed graphs. In addition, integrating multiple source domains can introduce domain conflicts and uneven relevance to the target task, ultimately hindering robust generalization. To overcome these challenges, we propose MDGPT , a text-free graph learning framework that unifies and transfers knowledge across different graph domains. We begin by introducing a set of domain tokens that align diverse features from various domains, enabling coherent representation learning during pre-training. Then, we develop a two-stage prompting mechanism: the unifying prompt encodes domain-invariant semantics shared across multiple sources, while the mixing prompt selectively composes domain-specific signals to adapt the model to the target domain. We evaluate our approach on six benchmark datasets, where MDGPT achieves up to a 10.2% improvement over state-of-the-art methods, demonstrating its effectiveness in cross-domain generalization. (Codes are available at https://github.com/Nashchou/MDGPT .)
Language: English
Open Access status: green
OpenAlex ID: W4399026295
FWCI: 29.78
Cites: 61
Cited by: 50
Related to: 10