Graph structure-aware causal graph rectification method based on LLMs

Causal graphs serve as essential tools for representing domain knowledge and play crucial roles in fault diagnosis, decision support, and other scenarios. However, initial causal graphs constructed by experts often suffer from missing relationships and incorrect connections, which severely impact the performance of downstream reasoning tasks. This paper proposes a Graph Structure-aware Causal Graph Rectification method (GSCGR) that combines the semantic understanding capabilities of LLMs with topological feature analysis of graph structures to achieve precise optimization of causal graphs. The method first identifies structurally critical edges through edge betweenness centrality, then constructs structured prompts to guide LLMs in generating rectification candidates, and finally completes rectification decisions based on multi-objective optimization of model confidence and structural improvement. Experiments in industrial fault diagnosis scenarios demonstrate that our method achieves significant performance improvements compared to traditional graph neural network approaches. This research provides a new technical pathway for the automated refinement of domain causal knowledge.