Reverse-graph enhanced graph neural networks for session-based recommendation
Reverse-graph enhanced graph neural networks for session-based recommendation
复制标题
DOI：
10.1016/j.eswa.2023.122995
复制
发表时间：
2023-12
期刊：
影响因子：
--
通讯作者：
Hao Xu;Bo Yang;Xiangkun Liu
中科院分区：
文献类型：
--
作者：
Hao Xu;Bo Yang;Xiangkun Liu
Session-based recommendation (SBR) aims to predict the next item based on anonymous sessions, which has become a hot research area in recent years. A variety of SBR models have been proposed, and the graph neural network (GNN)-based models are shown to achieve the state-of-the-art performance in SBR. In our research, we observe that there is a decrease in the performance of SBR models for sessions with short length, which hinders the overall performance of SBR models. To alleviate this problem, we propose to add some virtual item(s) into short sessions as the supplementary contextual information to improve model performance. The main challenge lies in how to predict the appropriate virtual item(s) to be added in short sessions. In this paper, we propose Reverse-graph Enhanced Graph Neural Networks (RGE-GNN), a new GNN-based model for SBR. In RGE-GNN, we first construct a reverse graph of a session (RGS) to capture the complex reverse transition relation between items in the session. After that, the reverse prediction is proposed to be conducted to predict the appropriate virtual item(s). As a result, richer information could be carried by the enhanced session, which could lead to higher recommendation accuracy. In addition, to ensure that more relevant virtual item(s) should be added into a session, we propose a preference memory mechanism (PMM) to obtain the expected preference of each session. Extensive experiments conducted on three real-world datasets demonstrate that RGE-GNN outperforms the baselines including the recent SBR models.
