关系生成图注意力网络的知识图谱链接预测

Knowledge graph link prediction based on relational generative graph attention network

针对实体邻域三元组缺少联系的问题,提出基于关系生成图注意力网络(RGGAT)的知识图谱链接预测方法.利用不同类型的关系生成相应的注意力机制参数,邻域三元组按照关系类型使用对应的参数计算注意力系数.实体通过聚合以关系为主导的邻域三元组信息得到更丰富的嵌入向量.在训练过程中对编码器和解码器进行共同训练,将编码器更新的实体向量和关系向量直接输入到解码器中,保证编码器和解码器训练目标一致.在3个公开数据集上进行链接预测实验,对比实验选用目前主流的5个模型作为基线.RGGAT方法在3个数据集上的Hits@10能达到0.5198、0.5104和0.9739,高于传统图注意力网络嵌入方法的.在邻域聚合阶数对比实验中,1阶关系邻域聚合的方法相比2阶关系在Hits@10上提升3.59%.

A knowledge graph link prediction method for relational generative graph attention network(RGGAT)was proposed to address the problem of missing links in entity neighborhood triples.Different types of relation were used to generate the corresponding attention mechanism parameters,and the attention coefficient was calculated by the neighborhood triples through the corresponding parameters according to the relation types.The entity got a richer embedding vector by aggregating the relation-dominated neighborhood triples information.The encoder and the decoder were jointly trained during the training process,and the entity vector and relation vector updated by the encoder were directly input into the decoder to ensure that the training objectives of the encoder and the decoder were consistent.The link prediction experiment was carried out on three public datasets,and five current mainstream models were selected as the baseline for the comparison experiment.The Hits@10 of RGGAT method on the three datasets were 0.5198,0.5104 and 0.9739,higher than that of the traditional graph attention network embedding method.In the comparison experiment of neighborhood aggregation order,the Hits@10 of the neighborhood aggregation method for one-hop relation was improved by 3.59%compared with the method for two-hop relation.