高级检索
    储晓恺, 范鑫鑫, 毕经平. 基于K阶互信息估计的位置感知网络表征学习[J]. 计算机研究与发展, 2021, 58(8): 1612-1623. DOI: 10.7544/issn1000-1239.2021.20210321
    引用本文: 储晓恺, 范鑫鑫, 毕经平. 基于K阶互信息估计的位置感知网络表征学习[J]. 计算机研究与发展, 2021, 58(8): 1612-1623. DOI: 10.7544/issn1000-1239.2021.20210321
    Chu Xiaokai, Fan Xinxin, Bi Jingping. Position-Aware Network Representation Learning via K-Step Mutual Information Estimation[J]. Journal of Computer Research and Development, 2021, 58(8): 1612-1623. DOI: 10.7544/issn1000-1239.2021.20210321
    Citation: Chu Xiaokai, Fan Xinxin, Bi Jingping. Position-Aware Network Representation Learning via K-Step Mutual Information Estimation[J]. Journal of Computer Research and Development, 2021, 58(8): 1612-1623. DOI: 10.7544/issn1000-1239.2021.20210321

    基于K阶互信息估计的位置感知网络表征学习

    Position-Aware Network Representation Learning via K-Step Mutual Information Estimation

    • 摘要: 随着网络结构数据持续、快速的增长,各种复杂网络数据分析与应用层出不穷.近年来,网络表征学习已经成为各类网络分析任务的主流方法.网络表征学习的主要目标是依据节点间连接关系,学习高质量的节点表征向量,从而辅助分析下游任务.然而,现有的表征学习方法未考虑节点在网络中的位置信息.为了解决这一问题,提出了一种位置感知网络表征学习模型PMI,该模型通过最大化每个中心节点与各阶邻居之间的互信息,从而将节点的位置信息学入表征向量中.在表征训练过程中,PMI模型激励每个中心节点记住并识别其每阶的邻居节点,从而间接记录其位置信息.在4个不同领域的真实数据集上进行了多标签分类、网络重构、链接预测等多个代表性网络分析任务实验,实验结果表明提出的PMI模型可以学到高质量的节点表征向量,与现有的表征学习模型相比,PMI模型能够在多个下游任务上有较大幅度提升.此外,还设计邻居对齐任务对PMI模型进行进一步的分析,结果表明PMI模型学到的节点表征能够有效识别不同阶的邻居节点并捕获自身的位置信息,从而为各种下游任务生成合理有效的表征.

       

      Abstract: As the network data grows rapidly and persistently, also affiliated with more sophisticated applications, the network representation learning, which aims to learn the high-quality embedding vectors, has become the popular methodology to perform various network analysis tasks. However, the existing representation learning methods have little power in capturing the positional/locational information of the node. To handle the problem, this paper proposes a novel position-aware network representation learning model by figuring out center-rings mutual information estimation to plant the node’s global position into the embedding, PMI for short. The proposed PMI encourages each node to respectively perceive its K-step neighbors via the maximization of mutual information between this node and its step-specific neighbors. The extensive experiments using four real-world datasets on several representative tasks demonstrate that PMI can learn high-quality embeddings and achieve the best performance compared with other state-of-the-art models. Furthermore, a novel neighbor alignment experiment is additionally provided to verify that the learned embedding can identify its K-step neighbors and capture the positional information indeed to generate appropriate embeddings for various downstream tasks.

       

    /

    返回文章
    返回