针对现有的基于知识图谱的推荐算法往往侧重于物品端邻居信息,而忽视用户端兴趣特征问题,提出一种融合图注意力和知识图卷积网络的双端邻居推荐算法.首先,在用户端,以用户的历史兴趣作为种子,在知识图中迭代传播偏好,融合图注意力形成用户潜在兴趣向量;其次,在物品端,结合图卷积网络在知识图遍历路径中聚合重要邻域信息,获得物品偏好聚合向量;同时在损失函数中融入标签平滑正则化项;最后使用内积运算得到用户对物品的喜好预测.通过在公开数据集下的实验结果表明,文章算法与其他基准算法相比,在CTR(Click Through Rate)和Top-K(对模型给出的前K个预测结果进行性能评估)推荐场景下的评估指标AUC(Area Under Curve)、F1(F1-score)、recall(召回率)均有所提高.文章该算法具有较好的推荐性能和可解释性.
A Dual End Neighbor Recommendation Algorithm Integrating Graph Attention and Knowledge Graph Convolutional Networks
In order to solve the problem that the existing knowledge graph-based recommendation algorithms tend to focus on the object-side neighbor information and ignore the client-side interest features,a two-end neighbor recommendation algorithm based on the convolution network of graph attention and knowledge graph is proposed.Firstly,at the user side,the user's historical interests is taken as seeds,and the prefer-ences iteratively propagated in the knowledge graph,and the graph attention is fused to form the user's po-tential interest vector.Secondly,at the item side,combining graph convolutional network to aggregate impor-tant domain information in the knowledge graph traversal path to obtain the preference aggregation vector of the item.At the same time,a label smoothing regularization term is incorporated into the loss function.Final-ly,inner product operation is used to obtain the user's preference prediction for the item.The experimental results on public datasets show that,compared with other benchmark algorithms,this algorithm has improved evaluation metrics AUC(Area Under Curve),F1(F1-score),recall(recall rate)in CTR(Click Through Rate)and Top-K(Perform performance evaluation on the top K prediction results given by the model)recommenda-tion scenarios.Therefore,the proposed algorithm has good recommendation performance and interpretability.
NETL
NSTL
万方数据
