Abstract: To improve the accuracy of node importance evaluation in opportunistic networks, a node embedding-based node importance evaluation method is proposed. Considering the time-varying nature of opportunistic networks, time window aggregation graph is employed to represent the networks, so that network topology and temporal connections can be obtained in the window. Graph attention mechanism is utilized to extract the topological features of nodes, obtaining the topological embedding representation. Additionally, temporal embedding representation is obtained by using temporal encoding and self-attention mechanism to extract node temporal features. Node embedding vectors are achieved by integrating two representations. To reflect the information interaction between nodes, the cluster importance for nodes is introduced. The transition probability matrix is constructed, and the node importance is obtained by combining the PageRank algorithm. On three real opportunistic network datasets, experimental results demonstrate that the proposed method exhibits better evaluation accuracy compared to similar approaches such as f-PageRank and dynamic graph convolutional network (DGCN).

Key words: Opportunistic Network, Node Importance, Node Embedding, Self-Attention Mechanism, Transition Probability Matrix