基于生成对抗模型的异质信息网络语义表征方法研究  

Generative Adversarial Network Based Semantic Representation Learning for Heterogeneous Information Network

作　　者：赵瑜 谭海宁 刘志方 武超 ZHAO Yu;TAN Haining;LIU Zhifang;WU Chao(91977Troops,Beijing 100142,China;University of Chinese Academy of Sciences,Beijing 100049,China;Institute of Computing Technology,Chinese Academy of Sciences,Beijing 100190,China;Southwest Institute of Electronic Telecommunications Technology,Chengdu,Sichuan 610041,China;China Academic of Electronics and Information Technology,Beijing 100041,China)

机构地区：[1]91977部队,北京100142 [2]中国科学院大学,北京100049 [3]中国科学院计算技术研究所,北京100190 [4]西南电子电信技术研究所,四川成都610041 [5]中国电子科技集团公司电子科学研究院,北京100041

出　　处：《中文信息学报》2019年第11期83-94,共12页Journal of Chinese Information Processing

基　　金：国家重点研发计划(2017YFC0820700);国家自然科学基金(61472403)

摘　　要：近些年,网络表示学习问题吸引了大量研究者的关注,而异构信息网络由于其丰富的结构语义信息及其广阔的应用领域,更是成为了网络表示学习领域的重中之重。目前面向异构信息网络的表示学习模型主要可以分为基于生成式模型的表示学习方法和基于判别式模型的表示学习方法,但是很少有工作同时结合两种模型进行表示学习的优化。该文提出了结合生成式模型和判别式模型的异构信息网络表示学习模型HINGAN,主要是将对抗生成思想融入异构信息网络表示学习过程中,达到优化网络表示结果的目的。该模型首先在元路径的引导下构建带权信息网络图,然后在带权图上计算更新构造的生成器和判别器参数,通过生成对抗的博弈思想来获取最大收益。在AMiner和DBLP两个真实学术图谱数据集上的实验结果表明,HINGAN在多标签分类、链路预测以及可视化方面都能比现在主流的网络表示方法取得更优的效果,并且HINGAN可以应用于大规模的异构网络数据的表示和计算。除此之外,该文还总结了已有研究成果并对未来研究可能面临的挑战进行了展望。Due to the abundant structural and semantic information in the heterogeneous information network as well as its wide application,network representation learning for heterogeneous information networks has become a vital research issue.The current representation learning models for heterogeneous information network can be divided into generative model based or discriminative model based methods.In this paper,we propose a representation learning model for the heterogeneous information network called HINGAN,which integrates the generative adversarial network into the representation learning process of heterogeneous information network to improve network representational outcomes.Firstly,this model builds a weighted homogeneous information network in the guidance of the meta-path.Then,by employing the GAN for the greatest gain,it updates parameters of the constructed generator and discriminator on the weighted network.According to the experimental results on AMiner and DBLP,HINGAN can get a more outstanding effect than present mainstream network representation methods from the aspects of multi-label classification and visualization.At the same time,HINGAN can be applied to extensively scalable representation and effective calculation of the heterogeneous network data.

关 键 词：异质信息网络 语义信息挖掘 生成对抗网络 语义关系预测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]