.png){kind=icon}
.png){kind=icon}
Return
JGURD: joint gradient update relational direction-enhanced method for knowledge graph completion
.png){kind=icon}
OA
AI
DOI：10.7717/peerj-cs.2808.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Relational direction plays an important role in multi-relational knowledge graphs (KGs). Current knowledge graph completion (KGC) methods suffer from insufficient utilization of relation correlation information. To address this issue, this article proposes a novel KGC framework, namely JGURD, which uses the encoder-decoder structure to achieve Joint Gradient Update with Relational Direction information. It combines graph convolutional networks (GCNs) with KG embedding methods, defining a update mechanism for entities and relationships to joint gradient updates. To incorporate entity information into the update of relationships, the forward propagation gradients of the triple score function are recorded, and entity gradient information is fused into relationship updates. To fully utilize relational direction information, a relation correlation graph (RCG) is constructed based on the topological patterns of relationship pairs. We design a multi-relation encoder combining GCN and multi-layer attention mechanism on RCG to comprehensively capture local and global structures of the RCG. To enhance the interpretability and adaptability of JGURD, three different decoders are employed. Experimental results show that JGURD outperforms the second-place HHAN-KGC, and the Hits@3 and MRR metrics on the FB15k dataset increased by 6.8% and 8.9%, respectively.
Keywords:
Knowledge graph completion
Graph neural networks
Joint gradient update
Link prediction
Relational direction
Encoder-decoder
Multi-relational graph
Journal
IF:
2.5
Papers:
3.4K
Citations:
6.9K