Abstract: Addressing the insufficient exploration of project conversion information within the current session recommendation and the limited utilization of other session details, a multi-layer information interactive fusion algorithm based on graph neural network was proposed for session-based recommendation. Based on the current session, firstly, the information of neighborhood nodes was aggregated by designing different weights for the connection relationship between nodes, and the explicit information of project transformation in the current session is mined. Secondly, the neighborhood node information was aggregated by stacked residual graph attention network to mine the tacit information of item transformation in the current session. Finally, the sequence-dependent information present in the timestamp-based session was mined through a single-gated graph neural network. Based on other sessions, linking the entire set of sessions through the first-order neighbors of nodes, then, the four levels of embedding representation were integrated to obtain more comprehensive project conversion information, and the soft attention mechanism and reverse position embedding information were used to fuse the obtained project conversion information more effectively. Experimental results show that the P@20 and MRR@20 on diginetica dataset are increased by 0.98% and 0.84% respectively compared with the suboptimal model GCE-GNN, the P@20 and MRR@20 on the Tmall dataset are increased by 8.23% and 7.86% compared with the optimal model HyperS2Rec, and the P@20 and MRR@20 on the Nowplaying dataset are increased by 1.33% and 2.03% compared with the optimal model HyperS2Rec. 

Key words: session-based recommendation, residual graph attention network, gated graph neural network, soft attention, reverse position embedding 

摘要： 针对当前会话推荐中存在对于当前会话项目转换信息挖掘不充分且极少利用其他会话信息的问题，提出一种基于图神经网络的多层信息交互融合算法用于会话推荐。基于当前会话，首先，通过对节点之间的连接关系设计不同的权重聚合邻域节点的信息，挖掘当前会话中项目转换的显性信息；其次，通过基于堆叠的残差图注意力网络聚合邻域节点信息，挖掘当前会话中项目转换的隐性信息；最后，通过单门控图神经网络挖掘基于时间戳的会话中存在的序列依赖信息。基于其他会话，通过节点的一阶邻居将整个会话集联系起来，学习全局信息编码，进而融合四个层次的嵌入表示获得更全面的项目转换信息，同时使用软注意力机制和反向位置嵌入信息对于获得的项目转换信息进行更有效地融合。实验结果表明， 在diginetica数据集上P@20和MRR@20较次优模型GCE-GNN分别提升了0.98%、0.84%，在Tmall数据集上P@20和MRR@20较次优模型HyperS2Rec提升了8.23%、7.86%，在Nowplaying数据集上P@20和MRR@20较次优模型HyperS2Rec提升了1.33%、2.03%。

关键词: 会话推荐, 残差图注意力网络, 门控图神经网络, 软注意力, 反向位置嵌入