用于推荐系统的图卷积交叉网络

Graph Convolutional Cross Network for Recommender Systems

推荐系统中,因子分解机(FM)等特征交叉模型通常孤立地对待每个用户-物品交互样本,无法显式地利用样本中对象之间的隐含关系,造成了信息孤岛问题,导致学到的特征嵌入不是是最优的、冷门物品无法获得精准的推荐。因此,论文提出结合图表示学习和特征交叉的图卷积交叉网络(GraphCross):图卷积部分利用不同训练样本中对象的关联性构建异构图,并在此基础之上进行图卷积,使得生成的对象嵌入囊括其紧密相关的邻域节点对象的信息,破除了样本的孤立状态;特征交叉部分为FM模型,利用图卷积网络生成的对象嵌入构建特征交叉。GraphCross亦可推广为基于图表示学习-特征交叉的推荐算法框架。实验结果表明,利用图结构可有效提升推荐系统性能,尤其是针对冷门物品的推荐。

In recommender systems,feature cross models like factorization machine(FM)usually treat each user-item inter⁃action in isolation and cannot explicitly make use of relationship between objects,which causes the isolated information island prob⁃lem,resulting in suboptimal feature embeddings and hurting the recommendation accuracy for unpopular items.Therefore,the pro⁃posed graph convolutional cross network(GraphCross)combines graph representation learning together with feature cross.The graph convolution part builds a heterogeneous graph based on the relationship between objects,then operates graph convolution to produce object embeddings that aggregate information from neighbors,which breaks the isolation.The cross part is a FM model,fea⁃ture cross is constructed by utilizing the graph embeddings.GraphCross can also be generalized to a RS framework that consists of graph representation learnig and feature cross.Experiments demonstrate that making use of graph structures leads to a boost of the performance of RS,especially for unpopular items.