Graph machine learning (or graph model), represented by graph neural networks, employs machine learning (especially deep learning) to graph data and is an important research direction in the artificial intelligence field. With the emergence of the large language model (LLM), many research directions, such as Natural Language Processing (NLP) and Computer Vision (CV), have been significantly impacted. As for how LLM affects the graph model, a research team from Beijing University of Posts and Telecommunications published their new research on 15 December 2024 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature.
They think LLM is not able to effectively solve graph tasks, while the graph model does not possess the capabilities of LLM, e.g., emergence and homogenization. Inspired by the success of LLMs as foundation models in NLP, they explore the possibilities of graph foundation models (GFM) towards gaining emergence and homogenization capabilities. they analyze the characteristics of GFM and compare it with LLM, also pointing out its challenges and potential applications. It is worth notice that, this publish is defined as a high-level opinion piece, it has not delved into the details of technical implementations and experiments.

Fig 1 The illustration of Graph Foundation Model
In the research, they proposed the concept of Graph Foundation Model (GFM), encapsulating the vision of harmoniously merging the capabilities of graph models and large language models. As illustrated in Fig 1, a graph foundation model is envisioned as a model pre-trained on extensive graph data, primed for adaptation across diverse downstream graph tasks. Once developed, such models would offer increased expressive capabilities, better transferability, and suitability for complex graph tasks and data. They might even demonstrate emergence and homogenization capabilities on graph data, similar to how LLMs perform with natural language data. These benefits, highly desired by researchers in the field of GNNs, have the potential to establish a new leading approach in graph data processing.
DOI: 10.1007/s11704-024-40046-0