Abstract
Heterogeneous Information Networks (HINs) are complex structures hosting various types of entity nodes and diverse relationships. Meta-paths are crucial in HINs for relaying detailed information and enhancing knowledge transfer. However, their discovery is challenged by the expansive meta-path space inherent in complex HINs. To address this, we introduce LLM4HIN, an innovative meta-path discovery framework that utilizes Large Language Models (LLMs) for efficient meta-path discovery. Through the use of LLMs, LLM4HIN bypasses challenges including computational limitations, semantic inconsistency, and path selection optimization traditionally associated with meta-path reasoning. LLM4HIN could efficiently discovers meta-paths by utilizing LLMs to generate these paths from a minimal subset of HIN samples. Experimental evaluation on large-scale HINs attests to the enhanced performance of the meta-paths generated by LLM4HIN in link prediction tasks compared to state-of-the-art methods.