融合双端注意力网络的知识图谱推荐算法

Knowledge Graph Recommendation Algorithm Integrating Double-End Attention Network

知识图谱在推荐算法中的应用有效增强了推荐结果的可解释性,但仍缺乏对用户-物品交互信息的关注。针对目前已有的知识图谱推荐算法忽略用户端与物品端的差异而无法有效地提取用户与物品信息的问题,提出了融合双端注意力网络的知识图谱推荐算法(double end knowledge graph attention network,DEKGAN)。首先根据用户端与物品端两者不同的需求设计两种注意力网络,在用户端是将用户的浏览记录输入到知识图谱中不断扩展用户的潜在兴趣,通过基于用户的注意力网络对三元组内部信息进行计算以更准确地获取用户的嵌入表示向量;在物品端是将物品信息输入到知识图谱内中不断获取与用户喜好有关的物品信息,使用基于用户偏好的注意力网络指定更准确的传播方向,然后将两端注意力网络中获取到的邻域信息聚合生成物品的嵌入表示向量,最终获取用户对物品的喜好概率。通过在数据集MovieLens-1M和Book-Crossing中进行实验,采用AUC、F1、precision和recall指标进行评估,结果证明与其他基准算法相比推荐的准确性和可解释性有显著的提升。

The application of knowledge graphs in recommendation algorithms effectively enhances the interpretability of recommendation results,but still lacks attention to user-item interaction information.Aiming at the problem that the existing knowledge graph recommendation algorithms ignore the difference between the user end and the item end and cannot effectively extract user and item information,a knowledge graph recommendation algorithm integrating double-end atten-tion network(double end knowledge graph attention network,DEKGAN)is proposed.Firstly,two attention networks are designed according to the different needs of the user end and the item end.On the user side,the browsing records of user are input into the knowledge graph to continuously expand the potential interest of user,and the internal information of triples is calculated by user-based attention network to obtain the embedded representation vector of user more accurately.On the item side,the item information is input into the knowledge graph to continuously obtain the item information related to the preference of user,and the attention network based on the preference of user is used to specify a more accurate propagation direction.Then,the neighborhood information obtained in the attention networks at both ends is aggregated to generate the embedding representation vector of the item,and finally the preference of user probability for the item is obtained.By conducting experiments on the datasets MovieLens-1M and Book-Crossing,using AUC,F1,precision and recall metrics for evaluation,the results demonstrate a significant improvement in recommendation accuracy and interpret-ability compared to other benchmark algorithms.