Inductive Knowledge Graph Reasoning Algorithm Based on Semantic Bridging and Gated Aggregation

Knowledge graph reasoning is of great significance in scenarios such as intelligent question answering, recommendation systems, and semantic search. Path-based graph neural network (GNN) reasoning methods can uncover deep semantic associations by aggregating neighbor information along multi-hop paths. However, capturing semantic dependencies between distant nodes under a limited number of propagation steps remains a challenge. Concurrently, as network depth increases, the message aggregation process is prone to vanishing gradients and information dilution, further weakening the reasoning capability over long-range nodes or relations. To address these challenges, this paper proposes a novel inductive knowledge graph reasoning model named ReGaterNet. The model introduces a semantic relevance bridging (SRB) module. This module pre-selects starting entities that are highly relevant to the query. It then connects these entities to the head entity via high-speed shortcut channels. As a result, remote dependency information can be transmitted efficiently within a limited number of hops. Residual gated recurrent unit (Residual-GRU) is incorporated into the message aggregation process of each layer. This mechanism fuses the historical hidden state with the newly aggregated messages. It helps mitigate the issues of gradient vanishing and information dilution in deep networks. Extensive experiments conducted on the FB15k-237 and WN18RR inductive datasets demonstrate that ReGaterNet achieves superior performance compared to existing representative inductive reasoning models.