How LLMs Are Reinventing Graph Machine Learning
· Shijie Wang, Jiani Huang, Zhikai Chen, Yu Song, Wenzhuo Tang, Haitao Mao, Wenqi Fan, Hui Liu, Xiaorui Liu, Dawei Yin, Qing Li
A new survey on arXiv explores how large language models and graph machine learning are starting to work together, and the results are pretty interesting.…
A new survey on arXiv explores how large language models and graph machine learning are starting to work together, and the results are pretty interesting. Graphs power everything from social networks to knowledge bases and molecular discovery, while LLMs have become the go to tools for language, vision, and recommendations. Now researchers are asking what happens when you combine them.
The survey, posted on April 23, 2024 and updated since then, breaks the relationship into two directions. First, LLMs can help improve graph learning itself. They can generate better graph features, reduce the need for labeled data, and tackle tricky problems like graph heterophily (where connected nodes are different from each other) and out of distribution generalization. That's a big deal because these challenges have been hard to crack with traditional methods.
On the flip side, graphs can make LLMs smarter. Knowledge graphs, in particular, are packed with structured, reliable facts. Feeding that kind of grounding into LLMs could help with their well known problems: hallucinations, lack of explainability, and the need for huge amounts of training data. The survey covers how graphs are being used to enhance both pre training and inference.
We're still in the early days, but this crossover feels like a natural next step. Researchers are already experimenting with applications in drug discovery, recommendation systems, and question answering. The survey gives a solid overview of what's been done and points to what might come next, like better few shot learning and more trustworthy AI. For anyone following graph ML or LLMs, this is worth keeping an eye on.