基于聚合多阶邻域信息的细化方法的多粒度网络表示学习

Multi-granularity Network Representation Learning Based on the Refinement Method Aggregating Multi-neighboring Information

多粒度网络表示学习方法因其在学习节点表示过程中可以保留网络的多粒度特征而受到越来越多的关注.该类方法主要经过粗化和细化两个阶段.现有的工作侧重于设计粗化策略以压缩网络规模获得网络的多粒度结构.但是如何保留这种多粒度结构,将粗粒度空间的节点表示细化回原始网络仍具有挑战.本文提出一种基于聚合多阶邻域信息的细化方法的多粒度网络表示学习方法NRAM(Network Refinement based on Aggregating Multi-neighboring information).首先,对于粗化阶段生成的多粒度网络,仅利用现有的网络表示学习方法学习最粗粒度网络的表示;然后将从粗粒度网络继承的节点表示和细粒度网络的结构信息相融合得到细粒度网络的初始嵌入;最后通过聚合节点多阶邻域信息的方式得到细粒度网络的节点表示,迭代该过程直到获得原始网络的节点向量.在3个公共数据集上节点分类的结果证明了NRAM的有效性.

Multi-granularity network representation learning has attracted more and more attention because it can retain the multi-granularity characteristics of the network in the process of learning node representation.This kind of method mainly goes through two stages:network coarsening and network refinement.The existing works focus on designing coarsening strategies to reduce the scale of the network to obtain the multi-granularity structure of the network.However,it is still a challenge to retain multi-granularity structure and refine the node representation from the coarse-grained space to the original space.In this paper,we propose a multi-granularity network representation learning method NRAM(Network Refinement based on Aggregating Multi-neighboring information).Firstly,for the multi-granularity network generated in the coarsening phase,we only use one of existing network representation learning methods to learn the representation of the coarsest granularity network.Then the node representation inherited from the coarse-grained network and the structural information of the fine-grained network are fused to obtain the initial embedding of the fine-grained network.Finally,the node representation of fine-grained network is obtained by aggregating the multi-neighboring information.We can obtain the node representation of original network by repeating this process.The results of node classification on three common datasets demonstrate the effectiveness of NRAM.