面向复杂网络的异构网络表示学习综述

A survey of heterogeneous network representation learning for complex networks

异构信息网络包含丰富的节点信息和链接信息,具有复杂异质性、高稀疏性、属性高维性等特性,这些特性给网络表示学习任务带来了巨大的挑战。异构网络表示学习通过在嵌入过程中将多样化的异质信息和结构信息进行有效融合,学习得到更有利于下游机器学习任务的低维特征向量。从异构网络表示学习方法的研究粒度出发,对近年的研究现状进行了比较全面的分析和讨论。首先探讨网络表示学习的产生动机,阐述了近年的异构网络表示学习的研究历程;然后对具有代表性的算法模型进行分类讨论,归纳其主要的研究内容和所使用的嵌入技巧。最后给出了未来工作中异构网络表示学习可能的研究方向和比较有价值的研究内容。

Heterogeneous information networks contain rich information about node and link,and have some characteristics,such as complex heterogeneity,high sparsity,high-dimensionality of attributes,etc,which brings huge challenges to network representation learning tasks.The heterogeneous network representation learning learns low-dimensional feature vectors that are more conducive to downstream machine learning tasks by effectively integrating diverse heterogeneous information and structural information in the embedding process.It conducts a relatively comprehensive analysis and discussion of the research status in recent years,starting from the research granularity of the heterogeneous network representation learning method.Firstly,the motivation of network representation learning and the research history of heterogeneous information network representation learning in recent years was discussed.Then some representative algorithm models were classified,followed by the summary of their main research contents and embedding skills.Finally,some possible directions and valuable contents of heterogeneous information network representation learning research in future work were listed.