Effective and Efficient Graph Foundation Model

Graph data has emerged as a central component in numerous real-world applications, spanning recommender systems, drug discovery, social networking, and traffic forecasting. While traditional and modern graph learning techniques—ranging from graph kernels to Graph Neural Networks (GNNs) and graph transformers—have achieved significant success, their task-specific nature and reliance on supervised learning limit their adaptability to new, unseen tasks. This rigidity becomes especially problematic in dynamic environments where retraining for every new task is costly and often infeasible. Inspired by the transformative impact of foundation models in natural language processing, this thesis explores the feasibility of developing a graph foundation model—a single, unified model capable of generalizing across diverse graph tasks and domains. Such a model would not only address conventional challenges like node classification and link prediction but also extend to emerging tasks involving information retrieval, reasoning, and structural querying over graphs. To realize this vision, the thesis identifies and tackles three key challenges: (1) domain heterogeneity in graph data representations, (2) task inflexibility stemming from human-designed prediction heads, and (3) the computational and structural complexity intrinsic to graphs. The thesis is structured around three main contributions: Chapter 2 introduces a unified model capable of learning across multiple graph domains; Chapter 3 expands on this foundation to build a more generalizable and versatile graph model; and Chapter 4 investigates the fundamental properties of graph learning to enhance the efficiency and scalability of the proposed models. Together, these contributions represent a step toward building general-purpose graph models that mirror the flexibility and transferability of language foundation models, laying the groundwork for more adaptive and scalable graph-based machine learning systems.