Democratizing Large Language Model-Based Graph Data Augmentation via Latent Knowledge Graphs

Proposed DemoGraph uses large language models for black-box context-driven graph data augmentation. It generates knowledge graphs through text prompts, integrating structural interactions into original graphs during training. A granularity-aware prompting strategy and instruction fine-tuning module control sparsity, enhancing predictive performance and interpretability, especially in electronic health records. The framework addresses open-world settings with proper sparsity control, democratizing access to extensive knowledge without model weights or source codes. Key contributions include dynamic merging, granularity-aware prompting, and sequential prompting with instruction fine-tuning. This method improves graph learning tasks, particularly in electronic health records, by maximizing contextual information utilization, leading to better predictive performance and interpretability. Extensive experiments validate its effectiveness over existing methods, demonstrating high scalability across datasets.