Www 2022 | rethinking the knowledge map completion of graph convolution network

Article address ：https://arxiv.org/abs/2202.05679

Code address ：https://github.com/MIRALab-USTC/GCN4KGC

Figure convolution network (GCNs) It is an effective method of modeling graph structure , It is completed in the knowledge map (KGC) It's becoming more and more popular in China . be based on GCNs Of KGC The model first uses GCNs Generate expression entity representation , Then use knowledge graph to embed (knowledge graph embedding, KGE) The model captures the interactions and relationships between entities . However , Many are based on GCN Of KGC The model failed to surpass the most advanced KGE Model , Despite the introduction of additional computational complexity . This phenomenon urges us to explore GCNs stay KGC The real role of . therefore , This paper is based on GCNs Of KGC Based on the model , Introduce variables to find GCNs stay KGC Key factors in . It's amazing , We observed from the experiment ,GCNs Graph structure modeling in KGC The performance of the model has no significant impact , This is contrary to what people generally know . contrary , The transformation of entity representation is responsible for performance improvement . On this basis , We propose a simple and effective framework , be known as LTEKGE, The framework will integrate the existing KGE The model is combined with the entity embedding of linear transformation . Experiments show that ,LTE-KGE The model is based on gcn Of KGC The method has similar performance improvements , At the same time, it has higher computational efficiency . These results suggest that , The existing GCNs stay KGC It's not necessary , The new is based on GCNs Of KGC The model should rely on more ablation studies to verify its effectiveness .