LLMs for graph learning
Researchers have delved into various strategies for incorporating LLMs into the graph learning process. Each method presents distinct benefits and potential uses. Let’s look at some of the key functions that LLMs can fulfill.
LLMs as enhancers
Traditional GNNs rely on the quality of initial node features, often with limited textual descriptions. LLMs, with their vast knowledge and language comprehension abilities, can bridge this gap. By enhancing these features, LLMs empower GNNs to capture intricate relationships and dynamics within the graph, leading to superior performance on tasks such as node classification or link prediction.
There are two primary methods for harnessing LLMs as enhancers. The first is feature-level enhancement, which can be achieved in various ways using LLMs:
- Synonyms and related concepts: The LLM goes beyond the surface level of the text description by recognizing synonyms and semantically related concepts. This...
