Abstract: The existing graph neural network sequence recommendation methods only focus on the temporal information of sequences without considering the time interval information of sequences, and only use single-task mode for sequence recommendation, ignoring the auxiliary tasks that can enhance the data and improve the generalization ability, which will lead to the dynamic interaction preferences of users and interaction time information cannot be captured clearly. To alleviate the impact of these problems, a dynamic graph neural network sequence recommendation algorithm with enhanced time awareness (TaDGSR) is proposed, which has two advantages. Firstly, the method constructs a sequence as a dynamic graph with temporal information, incorporates time interval information and adds a time-gated attention network module to enhance the utilization of temporal information while capturing the higher-order dynamic connections between sequences. Secondly, the method adopts the long short-term prediction task with temporal threshold segmentation as its auxiliary task to enhance the utilization of temporal interval information representation, so that the model can better capture the dynamic preferences of users at different interaction intervals, and finally improve the performance of the sequence recommendation task. Experiments on the Beauty dataset, Games dataset, and CDs dataset of Amazon.com show that：(1) the proposed method achieves average improvements of 4.47%, 4.37%, 2.72%, 1.16%, 4.61% and 3.97% in the Hit@10 and NDCG@10 metric values for the three datasets, respectively, compared to the current newer benchmark method; (2) adding time interval information can effectively improve the recommendation performance of dynamic graph neural networks, and the use of a time-gated attention module and a long- and short-term prediction assistance task can both bring positive improvements to the prediction performance.

Key words: sequential recommendation, graph neural network, neighborhood aggregation, time-aware, multi-tasking learning

摘要： 现有的图神经网络序列推荐方法只关注序列的时序信息而没考虑到序列的时间间隔信息，且只采用单任务的模式进行序列推荐，忽略了可以增强数据，提高泛化能力的辅助任务，这将导致用户的动态交互偏好以及交互时间信息无法被清晰地捕捉。为了缓解上述问题带来的影响，提出了增强时间感知的动态图神经网络序列推荐算法（time-aware enhancement dynamic graph neural networks for sequential recommendation，TaDGSR），具有两方面的优点。该方法将序列构造为一个具有时序信息的动态图，在此基础上融入时间间隔信息并增加时间门控注意力网络模块，以便在捕捉序列间的高阶动态连接的同时增强对时间信息的充分利用。该方法采用了时间阈值分割的长短期预估任务作为其辅助任务，加强时间间隔信息表征的利用，使模型更好地捕捉不同交互间隔的用户动态偏好，最终提升序列推荐任务性能。通过在Amazon电商的Beauty数据集、Games数据集以及CDs数据集上进行的实验，结果表明：（1）与目前较新的基准方法相比，所提方法在三个数据集的Hit@10和NDCG@10指标值上分别取得了4.47%、4.37%、2.72%、1.16%、4.61%、3.97%的平均提升；（2）增加时间间隔信息可以有效提高动态图神经网络的推荐性能，采用时间门控注意力模块以及长短期预估辅助任务对预测性能均能带来正面提升。

关键词: 序列推荐, 图神经网络, 邻域聚合, 时间感知, 多任务学习