RED: Rule Guided Prompt Engineering for Graph Data Imputation

Graph-based data imputation remains challenging due to the coexistence of complex topological dependencies and semantic heterogeneity. Traditional statistical and rule-based methods struggle to simultaneously capture topological structures and semantic contexts, leading to suboptimal imputations under structural constraints. To address this limitation, we propose RED, a rule-guided prompt engineering framework that performs context-aware graph data imputation by aligning graph structural dependencies with LLM-based semantic reasoning. The approach linearizes matched subgraphs containing missing nodes and attributes into interpretable pseudo-relational tables and couples them with semanticized GDD constraints, enabling the LLM to reason under explicit dependency signals rather than implicit graph parsing. Extensive experiments on relational and graph datasets show that RED consistently outperforms Holoclean, Derand, and GIG in precision and recall on multiple benchmarks. Iterative refinement experiments confirm convergence, while ablation studies validate the complementary roles of subgraph context, semantic constraints, and structured representation.