融合社交关系的图神经网络序列推荐模型研究

Research on the Graph Neural Network-based Sequential Recommendation Model with Social Relationship Fusion

针对在线社区中用户兴趣偏好快速变化的现象,为了快速定位用户的当前兴趣同时考虑动态的社交因素对用户决策的影响等问题,提出一种融合社交关系的图神经网络序列推荐的模型.首先利用门控循环单元对用户最近一次的会话作为短期兴趣进行建模,而对于用户的朋友则使用短期兴趣和长期兴趣串联融合来表示,其中短期兴趣使用朋友最近一次的会话进行建模,长期兴趣则是学习过的个体嵌入;然后通过用户的社交关系构建用户-朋友无向单元图并使用图注意力网络更新用户基于社交关系的表征;最后使用全连接层将用户的短期兴趣表示与用户基于社交关系的表示进行融合得到最终的用户表示并以此来进行项目的推荐.通过在三个与在线社区有关的数据集上的实验验证了社交关系能有效地提高序列推荐的准确性.该模型在召回率Recall@k与归一化累计增益NDCG两个评价指标上与其他模型相比有明显提升,当评估指标k值取20时,与DGRec模型相比,其Recall@20指标在三个数据集上分别提升了10.3%,5.7%,1.7%,NDCG在三个数据集上分别提升了6.85%,5.05%,2.4%.

In the light of a rapidly-changing preferences of users of online communities and for the purpose of quickly locating users’current interests while considering the impacts of dynamic social factors on users’decision-making,this paper proposes a graph neural network-based sequential recommendation model with social relationship fusion.Firstly,the user’s most recent session is modeled as short-term interest using a gated recurrent unit,while the user’s friends are represented by a concatenated fusion of short-term interest and long-term interest,where the short-term interest of the friends is modeled based on a recent session and the long-term interest is the learned individual embedding.Then,a user-friend undirected unit graph is constructed through the user’s social relationship and a graph attention network is used to update the user’s social relationship-based representation.Finally,a fully-connected layer is used to fuse the user’s short-term interest representation with the user’s social relationship-based representation to obtain the final user representation to conduct items recommendation.Experiments on three datasets related to online communities verify that social relationship can effectively improve the accuracy of sequential recommendation.When the value of the evaluation index k is taken as 20,compared with the DGRec model,the Recall@20 index is improved by 10.3%,5.7%and 1.7%on the three datasets,respectively,and the NDCG is increased by 6.85%,5.05%and 2.4%on the three datasets,respectively.