Abstract:Existing knowledge graph reasoning models suffer from two main limitations. First, they often treat entities in isolation, neglecting the global dependencies among entities and relationships within the knowledge graph (KG). Second, graph neural network (GNN) typically assume independence among neighboring nodes, ignoring potential interaction features between them. To address these issues, a knowledge graph reasoning model based on a graph interactive attention network, termed LS-InGAT, is proposed by integrating short-range and long-range graph interaction mechanisms. A convolutional neural network (CNN) is employed to capture long-range entities and relationships, and an attention-based interaction mechanism is used to aggregate node-relation information, thereby generating long-range feature representations of entities. Additionally, an efficient multi-perspective neighborhood interactive graph attention network is introduced to simultaneously capture interactions between entities and interactions among local relationships. These diverse interaction features are fused to generate short-range structural representations. Finally, the combined long-range and short-range features are fed into a decoder for triple scoring. Experimental results on two public datasets demonstrate that LS-InGAT outperforms the state-of-the-art DRR-GAT model, achieving MRR improvements of 1.4% and 7.7% on the FB15k-237 and WN18RR datasets, respectively.