基于同质子图变换的异质网络表示学习  

Heterogeneous Network Representation Learning Based on Homogeneous Subgraghs Transformation

作　　者：尹赢 吉立新[1] 程晓涛[1] 黄瑞阳[1] 刘正铭 YIN Ying;JI Lixin;CHENG Xiaotao;HUANG Ruiyang;LIU Zhengming(China National Digital Switching System Engineering&Technological R&D Center,Zhengzhou 450002,China)

机构地区：[1]国家数字交换系统工程技术研究中心

出　　处：《计算机工程》2019年第11期204-212,共9页Computer Engineering

基　　金：国家自然科学基金创新群体项目(61521003)

摘　　要：目前针对信息网络的研究多数基于同质网络,关于异质信息网络的网络表示学习研究相对较少。为此,提出一种结合不同元路径将异质信息网络转化成带权同质子图的网络表示学习算法。基于不同元路径在同类节点间构建带权重的连边,从异质网络中抽取出带权同质子图,通过带偏置的随机游走方式得到同类节点序列,并利用Skip-gram模型生成该类节点的表示向量。实验结果表明,与只考虑单一路径的算法相比,该算法处理节点分类、相似性搜索等数据挖掘任务时均能得到较好的效果。Currently,most studies on information network are based on homogenous networks,while the studies on network representation learning of heterogeneous networks are relatively rare.Therefore,we propose a network representation learning algorithm,which transforms heterogeneous information networks into weighted homomorphic subgraphs according to different meta-paths.First,we construct weighted edges base on different meta-paths between similar nodes,so as to extract the weighted homogenous subgraphs from heterogeneous network.Then,we obtain the similar node sequence by random walk with offset.Finally,we use Skimp-gram model to generate the representation vector of the nodes.Experimental results show that compared with other algorithms which only consider a single meta-path,the proposed algorithm has better effects on data mining tasks such as node classification and similarity search.

关 键 词：网络表示学习 异质网络 元路径 同质子图 随机游走 Skip-gram模型 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]