基于跨会话知识图谱的图注意力网络推荐方法

Graph Attention Network Recommendation Method Based on Cross-Session Knowledge Graph

基于会话的推荐方法旨在根据匿名用户行为序列预测下一个项目。然而,现有会话推荐方法多基于当前会话建模用户偏好,忽略了会话间蕴含的语义信息及知识图谱中丰富的实体和关系信息,无法有效缓解数据稀疏性的问题。提出一种基于跨会话信息与知识图谱的图注意力网络推荐方法。通过有效整合跨会话信息和知识图谱中的项目知识构建跨会话知识图谱,利用知识感知的注意力机制计算各邻居节点的重要性分数,以更新项目节点表示,采用门控循环单元和图注意力网络将每个会话表示为该会话的当前偏好和全局偏好的组合。在此基础上,将会话嵌入和项目嵌入拼接后输入到多层感知机,得到目标会话和候选项目的预测分数,从而实现会话推荐。实验结果表明,与GRU4REC、SR-GNN、FGNN等方法相比,该方法在KKBOX和JDATA两个真实数据集上的推荐命中率分别至少提高了8.23和2.41个百分点,能有效增强会话推荐性能。

The Session-Based Recommendation(SBR)method aims to predict the next item according to the sequence of anonymous user behaviors.However,most existing session recommendation methods typically model user preferences based on the current session,ignoring the semantic information between sessions and the rich relationship and entity information in the knowledge graph,which cannot effectively alleviate the data sparsity problem.This paper presents a recommendation method for a graph attention network based on the cross-session information and knowledge graph.The cross-session knowledge graph is constructed by effectively integrating cross-session information and project knowledge in the knowledge graph.The importance score of each neighbor node is calculated based on the attention mechanism of knowledge perception to update the project node representation.Each session is represented as a combination of the current and global preferences of the session using the Gated Recurrent Unit(GRU)and the graph attention network.Next,session embedding and item embedding are spliced and input into the Multi-Layer Perceptron(MLP)to obtain the prediction scores of target sessions and candidate items to achieve session recommendation.The experimental results show that compared with GRU4REC,SR-GNN,FGNN and other methods,the recommended hit rate of the proposed method on KKBOX and JDATA real data sets has increased by 8.23 and 2.41 percentage points respectively,which can effectively enhance the performance of session recommendation.