基于网络表示学习的机会网络链路预测  

Link Prediction in Opportunistic Networks Based on Network Representation Learning

作　　者：刘琳岚[1] 宋修洋 陈宇斌[2] LIU Linlan;SONG Xiuyang;CHEN Yubin(School of Information Engineering,Nanchang Hangkong University,Nanchang 330063,China;School of Software,Nanchang Hangkong University,Nanchang 330063,China)

机构地区：[1]南昌航空大学信息工程学院,南昌330063 [2]南昌航空大学软件学院,南昌330063

出　　处：《北京邮电大学学报》2022年第4期64-69,103,共7页Journal of Beijing University of Posts and Telecommunications

基　　金：国家自然科学基金项目(62062050,61962037)。

摘　　要：针对机会网络的多维链路属性和网络结构动态变化的特点,提出基于网络表示学习的链路预测方法。设置切片时长,将机会网络转化为网络快照序列,利用多维链路属性表示每个快照内的链路状态。采用网络表示学习方法聚合邻居节点的多维链路属性,并映射为低维的属性嵌入矩阵;采用基于注意力机制改进的循环神经网络学习网络拓扑随时间动态演化的规律,提取属性嵌入矩阵之间的时序特征;在输出层建立时序特征与链路状态之间的映射关系,实现下一时刻整网的链路预测。在Infocom-05和Hyccups等数据集上的实验结果表明,与现有同类方法相比,所提方法具有更高的预测精度。According to the characteristics of topology frequent changes and multi-dimensional attributes in opportunistic networks, a link prediction method based on network representation learning is proposed. The opportunistic network is transformed into snapshots by setting time slot. The link state of each snapshot is represented by multi-dimensional link attributes. Then, the network representation learning method is adopted to aggregate the multi-dimensional link attributes of neighbor nodes, which are mapped into a low-dimensional embedding matrix. The recurrent neural network improved based on the attention mechanism is employed to learn the laws of the evolution of network topology, and to extract the timing features between embedding matrices. Through the output layers, the mapping relationship between time serial characteristics and link-state is established to implement the link prediction for network at the next moment. The experimental results on mainstream datasets, such as Infocom-05 and Hyccups show that the proposed method achieves higher prediction accuracy compared with the existing link prediction methods.

关 键 词：机会网络 链路预测 网络表示学习 注意力机制 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]