Graph ML and LLMs Unite: Survey Reveals How They Boost Each Other
New research shows LLMs can fix graph data scarcity and graphs can reduce LLM hallucinations.
A new survey from Shijie Wang and 10 co-authors (accepted by ACM TIST) systematically reviews the bidirectional synergy between Graph Machine Learning and Large Language Models. On one side, LLMs can enhance graph features by generating high-quality embeddings and labels, reducing reliance on expensive manually labeled data. The paper specifically addresses how LLMs help with graph heterophily (when connected nodes differ) and out-of-distribution generalization—two major hurdles in graph learning. It also covers few-shot and transfer learning scenarios where LLMs bring their broad language understanding to mitigate sparse graph data.
On the flip side, knowledge graphs—rich in factual, structured knowledge—are shown to boost LLMs during both pre-training and inference. Graphs provide grounding that helps reduce hallucinations and improve explainability. The survey highlights applications in recommender systems, drug discovery, and social network analysis, and identifies open challenges such as scalability and domain adaptation. This work serves as a roadmap for researchers and practitioners exploring graph-neural hybrid systems.
- Explores how LLMs improve graph feature quality and overcome data scarcity in graph ML
- Shows that knowledge graphs enhance LLM pre-training and inference, reducing hallucinations
- Addresses graph heterophily and out-of-distribution generalization using LLM capabilities
Why It Matters
Combining graph structure with language models unlocks smarter reasoning and more reliable AI systems.