Abstract
Knowledge graphs (KGs) support knowledge organization and reasoning throughout the home appliance manufacturing lifecycle, but missing relations limit their use in downstream applications. To improve relation completion under sparse data conditions, this study proposes the Subgraph Knowledge Relation Attention Completion Network with Contrastive Learning (SKRA-CL). Given a target entity pair, SKRA-CL extracts a local subgraph and models two complementary sources of topological relation information: neighborhood relations and relation paths. An attention-based relation messaging mechanism identifies informative neighborhood relations, while a gating mechanism adaptively balances neighborhood and path-level evidence for relation prediction. A structure-aware negative sampling strategy further constructs informative negative triples based on entity structural similarity and relation context similarity. The model is trained using a weighted sum of contrastive and cross-entropy losses. On AMKGD, SKRA-CL achieves an MRR of 0.921, Hits@1 of 0.857, and Hits@3 of 0.982, yielding the highest MRR and Hits@3 while matching the highest Hits@1 among the compared methods. On WN18RR, the corresponding scores are 0.957, 0.924, and 0.988, respectively. These results support the effectiveness of adaptive topological relation fusion and structure-aware optimization for relation completion in sparse manufacturing KGs.