基于时空卷积的机会网络拓扑预测

Topology prediction for opportunistic network based on spatiotemporal convolution

机会网络拓扑的高动态性导致其拓扑预测极具挑战。现有拓扑预测方法主要关注网络长期时空依赖,忽视了短期时空特征。综合考虑机会网络长短期时空依赖关系,提出一种基于动态时间规整算法与时空卷积的机会网络拓扑预测方法(DTW-STC)。基于动态时间规整算法确定切片时长,将机会网络切分为快照,用快照的链路状态矩阵表征其拓扑信息;采用时序卷积神经网络获取短期时序特征,结合网络变化构建时空图表征短期时空关系,利用图卷积运算提取网络的短期时空特征,经过多次卷积的堆叠,得到网络长短期时空特征;基于自编码器结构实现向量空间切换,预测下一时刻网络拓扑。3个真实机会网络数据集ITC、MIT以及Asturias-er上的实验结果表明,DTW-STC方法的预测性能优于基线方法。

The high dynamics of opportunistic network topology leads to the challenges of topology prediction.The existing research mainly focuses on the long-term spatiotemporal dependence of networks,ignoring the short-term spatiotemporal features.A topology prediction method for opportunistic network based on dynamic time warping algorithm and spatiotemporal convolution(DTW-STC)was proposed,which integrated long-short term spatiotemporal dependence in opportunistic network.The time slot was determined based on dynamic time warping algorithm,so that the opportunistic network was sliced into snapshots which topology was presented by link state matrix.Temporal convolution was employed to extract short-term temporal features.The spatiotemporal graph,representing the short-term spatiotemporal relationship,was constructed by temporal features and network changes.The short-term spatiotemporal features were captured by graph convolution.After stacks of spatiotemporal convolution,the long-short term spatiotemporal features of network were achieved.Based on the autoencoder structure,vector space transformation was realized,so that the future network topology was predicted.The results on three real opportunistic network datasets,ITC,MIT,and Asturias-er,show that the proposed DTW-STC has better prediction performance than ones of other baseline methods.