融合GAT和CapsNet的链接预测研究

Research on Link Prediction with Combination of GAT and CapsNet

为解决现有的链接预测模型不能有效考虑三元组之间潜在关系的局限性,提出了一种融合图注意力网络(Graph Attention Networks,GAT)和胶囊神经网络(Capsules Noural Networks,CapsNet)的知识图谱链接预测模型,使用图注意力捕获每个实体邻域中的实体和关系特征,引入胶囊神经网络来解码三元组,通过胶囊神经网络节点嵌入特征的学习,生成连续向量与权重向量做点积运算,再构建评分函数用于判断三元组的准确性。在WN18RR和FB15K-237数据集上进行实验,结果表明该模型可以有效处理链接预测任务。

In order to solve the limitation that existing link prediction models cannot effectively consider the potential relationship between triples,this paper proposes a link prediction model of knowledge graph that integrates GAT(Graph Attention Networks)and CapsNet(Capsules Neural Networks).It uses the feature embedding method of graph attention to capture the entity and relationship features in each entity’s neighborhood,and introduces the capsule neural network to decode triples.Through the learning of embedded features of two-layer capsule neural network nodes,continuous vector is generated to perform dot product with weight vector,and then a scoring function is constructed to judge the accuracy of the triad.Experiments on WN18RR and FB15K-237 datasets indicate that the proposed model is effective in processing link prediction tasks.