基于双端权重约束的异质超网络表示学习  

Heterogeneous hypernetwork representation learning with dual-end weight constraints

作　　者：陈毅艰 朱宇 王晓英[1,2] 黄建强 曹腾飞 王威 Chen Yijian;Zhu Yu;Wang Xiaoying;Huang Jianqiang;Cao Tengfei;Wang Wei(School of Computer Technology&Application,Qinghai University,Xining 810016,China;Qinghai Provincial Laboratory for Intelligent Computing&Application,Qinghai University,Xining 810016,China)

机构地区：[1]青海大学计算机技术与应用学院,西宁810016 [2]青海大学青海省智能计算与应用实验室,西宁810016

出　　处：《计算机应用研究》2025年第2期406-412,共7页Application Research of Computers

基　　金：国家自然科学基金资助项目(62166032,62162053);青海省自然科学基金资助项目(2022-ZJ-961Q)。

摘　　要：与传统网络不同,超网络具有复杂的高阶元组关系,而现有大多数超网络表示学习方法不能很好地捕获复杂的高阶元组关系。针对上述问题,为了更好地捕获复杂的高阶元组关系,提出了基于双端权重约束的异质超网络表示学习方法。首先,该方法提出一个超边多源随机游走融合算法,将超边融入到基于超路径的随机游走节点序列中;然后,受到知识表示学习模型TransE的启发,该方法引入超边感知器模型与hyper-gram模型进行加权融合,以便于捕获超网络中复杂的高阶元组关系;最后,在四个真实超网络数据集上的实验表明,对于链接预测任务,该方法的性能几乎优于所有基线方法。对于超网络重建任务,在GPS数据集上,该方法的性能优于所有基线方法;同时,在drug数据集上,在超边重建比例大于0.3时,该方法的性能优于所有基线方法。总之,所提方法能够有效地捕获超网络中复杂的高阶元组关系。Different from traditional networks,the hypernetworks possess complex higher-order tuple relationships,which fail to be captured by most existing hypernetwork representation learning methods effectively.To address this issue and better capture complex higher-order tuple relationships,this paper proposed a heterogeneous hypernetwork representation learning me-thod with dual-end weight constraints abbreviated as HRDC.Firstly,this method proposed a hyperedge multi-source random walk fusion algorithm that incorporated the hyperedges into random walk node sequences based on the hyperpaths.Secondly,inspired by the knowledge representation learning model TransE,this method introduced hyperedge perceptron model and hyper-gram model to carry out weighted fusion,so as to capture complex higher-order tuple relationships in the hypernetworks.Finally,the experiments on four real-world hypernetwork datasets demonstrate that for link prediction tasks,the performance of this method is almost superior to all baseline methods.As for hypernetwork reconstruction tasks,on the GPS dataset,the performance of this method surpasses all baseline methods.On the drug dataset,when the hyperedge reconstruction ratio exceeds 0.3,the performance of this method outperforms all baseline methods.In summary,the proposed method can effectively capture complex higher-order tuple relationships in the hypernetworks.

关 键 词：超网络表示学习 双端权重约束 超边感知器 链接预测 超网络重建 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]