基于多重注意力机制的社会化推荐算法研究

Research on Social Recommendation Algorithm Based on Multiple Attention Mechanism

图神经网络是近年来兴起的一种对图结构数据进行学习的深度神经网络模型,由于其出色的特征学习能力,被广泛应用于推荐系统研究中。本文提出一种基于多重注意力机制的图神经网络社会化推荐算法,将用户、项目和用户社交信息嵌入到同一异构图中,并运用多重注意力网络来分别学习各部分信息的特征融合权重。在信息聚合部分,使用注意力机制来获取不同邻居节点的重要性权重,并对邻居节点的初始特征进行聚合;在融合邻居节点特征与自身特征时,引入注意力机制以获得节点的最终特征;在获得最终特征后的评分预测阶段,引入注意力机制来获取项目与用户特征的权重。多重注意力网络的引入能够有效区分不同邻居之间重要性的差异,从而获得更为精准的特征融合。在Epinions与Filmtrust数据集上进行实验,结果表明本文提出的方法优于其他几种性能较为优良的推荐算法。

Graph neural network is a deep neural network model emerged in recent years for learning graph-structured data,and has been widely applied to recommender system study due to its excellent feature learning ability.In this paper,we propose a graph-neural-network-based social recommendation algorithm with multiple attention mechanism,which embeds the user,item and user social information into the same heterogeneous graph,and employs the multiple attention network to learn the feature fusion weights of information of each part.In the information aggregation part,an attention mechanism is used to get the importance weights of different neighbor nodes,and the initial features of neighbor nodes are aggregated.The attention mechanism is also introduced in fusing the neighbor node features with its own features to obtain the final features of the node.After obtaining the final features,the attention mechanism is introduced to obtain the weights of the item and user features in the rating prediction stage.The introduction of multiple attention networks can effectively distinguish the difference in importance between different neighbors,thus obtaining more accurate feature fusion.Experiment is conducted on the Epinions and Filmtrust datasets,and the results show that the proposed method in this paper outperforms several other advantageous recommendation algorithms.