Research on plant knowledge graph reasoning based on dual-channel attention and topological perception.

Knowledge graphs (KGs) in the plant domain frequently contain "long-range dependency" paths (e.g., taxonomic hierarchies or ecological association chains of 4 or more hops), which pose significant challenges to existing KG reasoning models. Traditional Graph Neural Network (GNN) models struggle to effectively capture such long-range dependencies due to issues of long-distance information compression and over-smoothing. To address this, we propose the KRGAI-PLANT model, an inductive reasoning framework specifically designed for plant knowledge graphs. This model features a dual-channel architecture that integrates a global attention mechanism with local topology perception, enabling synergistic learning between global semantic interactions and local structural features among plant entities. We conduct experimental validation on subsets constructed from mainstream plant knowledge graphs, including DBpedia and PlantNet-KG. The results demonstrate that KRGAI-PLANT achieves significant improvements over baseline models such as GraIL and NBFNet in key evaluation metrics, including Hits@10 and AUC-PR, particularly exhibiting strong advantages in handling long-path reasoning tasks. This study provides an effective reasoning tool for knowledge discovery and association prediction in the plant domain. Furthermore, the proposed KRGAI-PLANT model offers a robust cognitive framework that can empower distributed and autonomous agricultural systems by converting complex, multi-source plant data into actionable knowledge for precise decision-making.