基于卷积图神经网络的多粒度表示学习框架

Multi-granular representation learning framework for Convolutional Graph Neural Networks

卷积图神经网络(Convolutional Graph Neural Network,ConvGNN)以其强大的表达能力被广泛应用于社交网络、生物网络等领域的网络表示学习中,多粒度网络表示学习已被证明能够改善已有网络嵌入方法的性能,但目前尚缺乏以改善ConvGNN性能为目标的框架.针对此问题,提出一种基于ConvGNN的多粒度网络表示学习框架M-NRL,分为四个模块:粒化模块、训练模块、推理模块和融合模块.粒化模块构造从细到粗的多粒度网络并保留不同粒度节点的属性和标签信息,训练模块在最粗粒度的网络上以端到端的方法训练任意一种ConvGNN并优化其模型参数,推理模块使用优化后的ConvGNN推理出不同粒度网络的节点表示,融合模块采用注意力权重聚合不同粒度的节点表示以产生最终的节点表示.在四个公开引文网络数据集上进行的半监督节点分类任务验证了M-NRL的有效性,实验结果表明,MNRL不仅能加速现有ConvGNN的训练,还可以增强其最终的表示质量.

Due to its powerful representational capabilities,Convolutional Graph Neural Networks(ConvGNN)have been widely used for network representation learning in social networks,biological networks,and other domains.There is a lack of a framework to increase the performance of ConvGNN,even though multi-granular network representation learning has been demonstrated the improvement for current network embedding approaches.To address this problem,a ConvGNN-based multi-granular network representation learning framework,called M-NRL,is proposed,which is divided into four modules:granulation,training,inference and fusion modules.The task of the granulation module is to construct fine-to-coarse multigranular networks and retain the attribute and label information of the nodes at different granularity networks.The task of the training module is to train any kind of ConvGNN on the coarsest network in an end-to-end approach and optimize its model parameters.The task of the inference module is to use the optimized ConvGNN to reason about the node representations of the networks at different granularity networks.The task of the fusion module is to use attention weights to aggregate node representations of different granularities to produce the final node representation.The semi-supervised node classification task is carried out on four public citation network datasets to verify the effectiveness of M-NRL.Experimental results show that M-NRL accelerates the training of the existing ConvGNN models and enhances the quality of its final representation.