New AI Framework Uses LLMs and GNNs for Knowledge Graph Hypothesis Discovery
A new framework for enhancing knowledge graphs transitions from merely confirming existing relationships to discovering structured hypotheses. This method utilizes graph neural networks for identifying phenotypes and making causal inferences, alongside large language models for generating hypotheses. The expansion of the knowledge graph is approached as a multi-objective optimization challenge, assessing potential claims for relevance, structural integrity, and originality. The selection of Pareto-optimal claims ensures a balance between confirming existing knowledge and exploring less documented relationships. The framework emphasizes data-supported relationships that are inadequately covered in current literature. It merges probabilistic reasoning with LLM-based claim extraction in a cohesive process. Existing techniques mainly focus on known relationships, while this paper introduces an evidence-based approach for controlled expansion driven by phenotype analysis.
Key facts
- Framework shifts from confirmatory to hypothesis-driven knowledge graph construction
- Integrates graph neural networks for phenotype discovery and causal inference
- Uses large language models for hypothesis generation and claim extraction
- Formulates knowledge graph expansion as multi-objective optimization problem
- Evaluates candidate claims based on relevance, structural validation, and novelty
- Employs Pareto-optimal selection to identify non-dominated claims
- Prioritizes structurally supported but underexplored relationships
- Combines probabilistic reasoning with LLMs in unified pipeline
Entities
—