【摘要】 现有的知识图推荐模型通过聚合相邻实体节点的特征和结构信息来更新当前位置实体节点的嵌入表示,为了控制计算成本和维护模型的稳定性,通常使用随机的固定大小的采样邻域来替代完整的知识图.然而,这些方法存在两个问题:首先,随机选择的邻域限制了知识图用于辅助推荐的效果和稳定性.此外,多数模型只是对所采样邻居节点特征进行均值聚合,这种聚合方法没有充分挖掘所采样邻居节点对于目标节点影响的差异性.针对以上问题,本文提出了基于关系紧密度的重要性采样方法,通过计算关系紧密度选择对目标节点更重要的邻域,以及基于池化操作的聚合方法,通过引入池化层训练得到不同邻居节点对目标节点的差异化权值.在结合本文提出的两种方法后,本文提出基于图神经网络的知识图推荐算法KGCN-PL.最后,本文评估了所改进模型在5个真实世界数据集上的性能,与近几年提出的基于知识图的推荐算法进行对比,在AUC,召回率指标上均取得提升.更多还原

【Abstract】 The existing knowledge graph recommendation model updates the embedded representation of the current location entity node by aggregating the characteristics and structural information of adjacent entity nodes. In order to control the calculation cost and maintain the stability of the model,random fixed-size neighborhood sampling is usually used to Replace the complete knowledge graph. However,there are two problems with these methods: First,randomly selected neighborhoods limit the effectiveness and stability of the knowledge graph used to assist recommendation. In addition,most models only perform mean aggregation on the characteristics of the sampled neighbor nodes. This aggregation method does not fully exploit the difference in the influence of the sampled neighbor nodes on the target node. In viewof the above problems,this paper proposes an importance sampling method based on the relationship tightness to select the neighborhood that is more important to the target node by calculating the relationship tightness,and an aggregation method based on the pooling operation,and introduces the pooling layer training to obtain different neighbor nodes Differentiated weights for target nodes. After combining the two methods proposed in this paper,this paper proposes a knowledge graph recommendation algorithm KGCN-PL based on graph neural network. Finally,this paper evaluates the performance of the improved model on five real-world datasets. Compared with the recommendation algorithm based on knowledge graphs proposed in recent years,the AUC and recall rate indicators have been improved.更多还原