Democratizing large language model-based graph data augmentation via latent knowledge graphs.

Area of Science:

• Graph representation learning
• Artificial intelligence
• Data science

Background:

• Graph data augmentation is crucial due to data scarcity and noise.
• Existing methods often ignore contextual information, focusing only on graph structure.
• Current large language model (LLM)-based graph learning methods are often white-box, limiting accessibility.

Purpose of the Study:

• To propose a black-box, context-driven graph data augmentation method guided by LLMs.
• To leverage LLM-generated knowledge graphs (KGs) for capturing structural interactions from text.
• To enhance graph learning by integrating contextual information and improving accessibility.

Main Methods:

• Developed DemoGraph, a black-box approach using LLMs for context-driven graph augmentation.
• Utilized text prompts to guide LLMs in generating KGs, capturing structural interactions.
• Implemented a dynamic merging schema for stochastic integration of generated KGs.
• Introduced granularity-aware prompting and instruction fine-tuning to control augmented graph sparsity.

Main Results:

• DemoGraph demonstrated superior effectiveness compared to existing graph data augmentation methods.
• The approach showed significant improvements in graph learning tasks, particularly with electronic health records (EHRs).
• Enhanced predictive performance and interpretability were observed, validating the method's contextual knowledge utilization.

Conclusions:

• DemoGraph effectively addresses limitations of existing graph augmentation techniques by incorporating contextual information.
• The black-box, LLM-guided approach democratizes advanced graph learning.
• The method shows strong potential for applications requiring rich contextual understanding, such as EHR analysis.