融合MAML和TGAT的机会网络动态链路预测模型  

Opportunistic Networks Dynamic Link Prediction Model Integrating MAML and TGAT

作　　者：袁子轩 张峰 许岗[2] 魏光辉 石永强 YUAN Zixuan;ZHANG Feng;XU Gang;WEI Guanghui;SHI Yongqiang(School of Computer and Information Technology,Shanxi University,Taiyuan 030006,China;College of Computer Science,Inner Mongolia University,Huhhot 010021,China)

机构地区：[1]山西大学计算机与信息技术学院,太原030006 [2]内蒙古大学计算机学院,呼和浩特010021

出　　处：《小型微型计算机系统》2024年第12期2957-2963,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62061036,6207703)资助;山西省自然科学基金面上项目(201901D111035)资助。

摘　　要：机会网络中由于节点频繁移动导致节点间的间歇性连接,导致网络拓扑结构会随时间不断发生变化,而采用离散时间切片的机会网络链路预测方法不能很好地刻画机会网络链路动态变化的演变规律.本文提出一种基于元学习框架的模型(METATGAT)来预测动态网络链接的变化情况,通过具有分层时间间隔和节点自适应的元学习者来提取网络动态演变的先验知识,并在元学习任务中设计一种采样注意力计算方式,来提升模型中图神经网络节点表征的计算效率.为了评估所提方法的有效性,本文选择Infocom05,Hyccups, Infocom06这3个真实的机会网络数据集进行验证,运用AUC和AP评价指标对所提方法与传统算法进行对比,结果显示所提方法能够有效捕获机会网络链路变化的一般知识,并且在链路预测准确率方面相比于传统算法有一定的提升.Due to the frequent movement of nodes,the network topology is changed owing to the intermittent connections between nodes in the opportunistic network over time.Link prediction using discrete time slice method cannot effectively depict the evolution patterns of the dynamic changes in the opportunistic network.A meta-learning based framework(METATGAT)of dynamic network link prediction is proposed in this paper,which extracts prior knowledge of dynamic network evolution by a meta-learner with hierarchical time intervals and node adaptation.The computational efficiency of node representation of graph neural network in the model is improved by designing a sampling attention calculation method in meta-learning tasks.In order to evaluate the effectiveness of this method,three real opportunistic network datasets,Infocom05,Hyccups,and Infocom06,are selected for experiments with AUC and AP metric.The results show that the proposed method can capture general knowledge of link changes in opportunistic network,and link prediction accuracy is improved compared to the traditional algorithms.

关 键 词：机会网络 链路预测 图嵌入 元学习 小样本学习 

分 类 号：TP393[自动化与计算机技术—计算机应用技术]