计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (17): 142-149.DOI: 10.3778/j.issn.1002-8331.1912-0018

• 模式识别与人工智能 • 上一篇    下一篇

对抗样本训练图分类器进行模型推理质量评估

于千城,於志文,王柱   

  1. 1.西北工业大学 计算机学院,西安 710072
    2.陕西省嵌入式系统技术重点实验室(西北工业大学),西安 710072
  • 出版日期:2020-09-01 发布日期:2020-08-31

Estimating Model Posterior Inference Quality by Using Adversarial Samples to Train Graph Classifier

YU Qiancheng, YU Zhiwen, WANG Zhu   

  1. 1.School of Computer Science and Engineering, Northwestern Polytechnical University, Xi’an 710072, China
    2.Shaanxi Provincial Key Laboratory for Embedded System(Northwestern Polytechnical University), Xi’an 710072, China
  • Online:2020-09-01 Published:2020-08-31

摘要:

重叠社区发现是社交网络分析与挖掘中的一个重要研究问题,现有的大部分方法都要求采用人工方法预先设定社区个数[K],这样做存在很多问题。将无限潜特征模型推广应用到关系型数据,以非参数贝叶斯层次模型为框架建立带重叠社区结构的网络生成模型,就可以避免预先设定[K]的值。然而,关系型无限潜特征模型的后验参数推理结果是一个[N×K]列的0、1矩阵上的概率分布,如何对这种多变量结构参数进行后验推理结果总结和后验推理质量评估是一个挑战,因此提出了一种利用基于对抗样本训练图卷积神经网络的图分类器来帮助总结推理结果和估计推理质量的方法。

关键词: 重叠社区发现, 非参数贝叶斯模型, 关系型无限潜特征模型, 参数推理质量估计, 图卷积神经网络

Abstract:

Overlapping community detection is a key research problem in social network analysis and mining, most of the existing methods require to predefine the number of communities K manually, this will cause a lot of issues. By extending the infinite latent feature model to relational data and building a network generation model based on the nonparametric Bayesian hierarchy model, setting the value of K in advance can be avoided. However, posteriori inference of the relational infinite latent feature model is a probability distribution over a binary matrices. How to summarize the posteriori inference result and estimate the quality of the posteriori inference for such multivariable structural parameter is a challenge. Therefore, a method is proposed which using adversarial samples to train the graph classifier to help to summarize the inference result and to estimate the quality of the inference via the adversarial training graph convolutional neural network.

Key words: overlapping community detection, nonparametric Bayesian model, relational Infinite Latent Feature Model(rILFM), estimate posterior inference quality, graph convolutional neural network