基于自监督学习的图转移网络会话推荐算法

Self-supervised graph transition network for session-based recommendation

针对基于会话的推荐算法存在建模物品表示缺乏会话间物品协同信息的问题,提出一种基于自监督学习的图转移网络会话推荐算法(S-SGTN)。该算法首先根据所有会话序列组建协同会话图;其次将当前会话与协同会话图中目标物品的邻居节点表示输入双通道图转移网络中,并在网络训练过程中引入自监督学习模块,通过最大化物品全局和局部表示的互信息,作为推荐任务的辅助任务,以改进物品和会话的表示;最后根据生成的匿名用户会话表示预测下一个产生交互的物品。在公开数据集上的实验结果表明,本文的推荐模型在召回率和平均倒数排名指标上的表现优于其他相关方法。

Session-based recommendation predicts the next item given anonymous sequence of previous items consumed in the session.Most existing studies focus on modeling the current session with various neural architectures.Those ap-proaches overlook many collaborative signals from the other sessions.This paper proposes a novel method,namely self-supervised graph transition network for session-based recommendation,S-SGTN for brevity.Methodologically,first,an undirected global collaborative graph based on all sessions is generated;then,the neighbor items’embed-ding of the target item in the current and collaborative session is input into the dual-channel graph neural networks.During network training,a self-supervised learning model is employed to improve the representation of items and sessions by maximizing the mutual information of global and local item representations as an auxiliary task to the recommendation task.Finally,the next interaction will be predicted based on the anonymous session representa-tion.Experimental results demonstrate that the proposed models show competitive or state-of-the-art performance in terms of Recall@k and MRR@k on three real-world datasets.