基于多头图注意力机制的神经协同过滤推荐算法

Neural Collaborative Filtering Recommendation Algorithm Based on Multi-head Graph Attention

目前主流的基于用户—项目二分图表征学习的推荐系统主要采用度归一化或平均池化的方法作为图上的近邻消息聚合策略,来学习用户和项目的表征向量.但是这两种聚合操作忽略了不同相邻节点对目标节点的重要性不同,从而导致最终刻画的用户和项目的表征表示不够准确.为此,本文提出了一种基于多头图注意力机制的神经协同过滤推荐模型(MGAT4Rec)来显示的建模目标节点与邻居节点之间的亲和力.该模型采用图注意力机制来捕获不同相邻节点的重要性并降低噪声信息的干扰,实现了对近邻节点信息的可解释性聚合;在此基础上,为了学习到更丰富的节点表征,通过使用多头图注意力机制来学习节点在不同潜在空间下的表征,将不同空间下的表征进行融合得到最终节点的表征向量.在MovieLens-100K和Amazon两个公开的数据集上进行了对比实验,MGAT4Rec在Recall@10和NDCG@10两个性能指标上相较于基线模型均有所提升.

Currently,the mainstream recommendation systems based on user-item bipartite graph representation learning mainly use degree normalization or Mean-pooling method as the neighbor aggregation scheme on the graph to learn the representation vectors of users and items.However,such aggregation operation ignores the varying importance of adjacent nodes to the target node,which leads to inaccurate embedding representations of users and items.Therefore,a neural collaborative filtering recommendation model(MGAT4Rec)based on multi-head graph attention mechanism is proposed in this paper to explicitly model the affinity between the target node and its neighboring nodes.MGAT4Rec leverages graph attention mechanism to capture the importance of different adjacent nodes and reduce the impact of noise information,obtaining interpretable aggregation of neighbor node information;based on this,in order to learn richer node representations,the proposed model employs multi-head graph attention mechanism to learn node representations in different latent spaces,and integrates the representations in different spaces to obtain the final representation vector of nodes.Comparative experiments conducted on two public datasets,MovieLens-100K and Amazon,and MGAT4Rec outperformed baseline models in both Recall@10 and NDCG@10 evaluation metrics.