基于图卷积网络的双向协同过滤推荐算法

Bidirectional Collaborative Filtering Recommendation Algorithm Based on Graph Convolution Network

近年来,图神经网络被证明是针对图数据研究的一个有效工具,而在推荐系统中用户和商品之间的关系可以被视为一个二分图,则可以尝试将图神经网络的相关技术原理应用到推荐系统。目前图神经网络方法存在局限有:高阶节点在信息传播过程中会携带噪声或负面信息;深层网络面临过平滑问题,即:网络深度达到一定层数后,由于节点特征聚合过多高阶节点信息,导致不同类别不可区分,此时模型的学习能力不增反降;两个问题都对后续学习任务造成负面影响,其学习策略存在进一步优化的空间。针对这一问题,本文基于图卷积神经网络设计了一个新的推荐算法--双向协同过滤推荐算法,将原本的大规模图划分为多个子图进行融合学习,并使用注意力机制对多个节点表示进行组合优化。相比于现有模型,该模型能够得到更高的准确度和召回率。

In recent years,graph neural network has been proved to be an effective tool for graph data research,in the recommendation system the relationship between users and products can be regarded as a bipartite graph,then we can try to apply the related technical principles of graph neural network to the recommendation system.The current graph neural network method has limitations:High-level nodes will carry noise or negative information in the process of information dissemination;The deep network has faced a smoothing problem,Namely:after the network depth reaches a certain number of layers,because node features aggregate too much high-order node information,it makes diff erent categories indistinguishable,at this time the learning ability of the model does not increase but decreases;Both problems have a negative impact on follow-up learning tasks,there is room for further optimization of its learning strategies.For the problem,This paper designs a new recommendation algorithm based on graph convolution neural network--Bidirectional collaborative filtering recommendation algorithm,the original large-scale graph is divided into multiple subgraphs for fusion learning,the attention mechanism is used to optimize the combination of multiple node representations.Compared to existing models,the model can get higher accuracy and recall.