No-CapsE:A Quaternion Capsule Network Knowledge Graph Completion Model Based on Node Co-occurrence
Knowledge graph completion addresses sparse data by predicting and filling in missing facts within the knowledge graph.Several models used for knowledge graph embeddings,such as CapsE and QuatE,have achieved satisfactory link prediction results.However,the capability of the CapsE model to perform link prediction and data mining in complex spaces may be restricted by its data dimensions,resulting in insufficient depth of data mining.QuatE uses quaternions to construct a hypercomplex plane for logical rotation.However,this method is simple and cannot effectively construct complex relationships.Therefore,this study proposes an improved capsule network completion method,No-CapsE,to construct a capsule network in a hypercomplex plane.Specifically,the data are represented by quaternions and input into a quaternion convolutional network.The output feature vectors are used as inputs to the capsule network,and the accuracy of the triplet is evaluated using dot product operations and judged based on the score.This study proposes the concept of node co-occurrence,representing entities and relationships as nodes to enhance model training speed.Finally,link prediction experiments are conducted on the publicly available datasets FB15K-237 and WN18RR Ablation experiments are designed and conducted to explore further the performance and effectiveness of the model used in this study.The results of both experiments indicate that the No-CapsE knowledge graph completion is more effective and suitable for large-scale linkage prediction tasks.
国家科技期刊平台
NSTL
万方数据
