Link Prediction for Opportunistic Networks Based on Feature Similarity

Objective An opportunistic network is a self-organizing network in which communication arises from the chances of node movement encounters. It can be applied in various scenarios, such as wildlife tracking, underwater rescue expeditions, and network coverage in remote areas. Link prediction involves determining whether a link is missing between two nodes or predicting whether a link will exist in the future, and this study aims to predict future links. Unlike ordinary networks, an opportunistic network is characterized by sparse node connections, complex structures, and frequent topology changes, leading to challenges in link prediction. Most existing opportunistic network link prediction methods employ network embedding algorithms for link prediction, but these algorithms have poor prediction effectiveness in networks with short average paths, and they do not fully utilize the network’s local structure information. Therefore, this study proposes a link prediction model for opportunistic networks based on feature similarity. The similarity of nodes is defined by considering node embedding vectors and local node information to improve the link prediction effect in opportunistic networks. Methods The Graph2vec model was employed to extract features of each snapshot. The sample entropies of the network, varying with candidate slicing slots, were obtained based on the network sample entropy calculation method. The slicing slot was determined using TOPSIS (technique for order preference by similarity to an ideal solution), which scores the candidate slicing slots by considering sample entropies and the training time of the Graph2vec model. The opportunistic network was sliced to generate a sequence of discrete network snapshots. The node embedding model based on GraphSAGE was utilized to extract the latent features of nodes. Considering that the connection duration of the nodes in the current snapshot influences the latent feature similarity of the nodes, the latent feature similarity was calculated by combining the Pearson correlation coefficient and the connection duration. The potential feature of the nodes alone was insufficient for extracting the features of network evolution. Therefore, the local structure information of the nodes was also considered. The topological similarity between nodes was obtained by evaluating the degree of nodes, the distance between nodes, and the number of paths between nodes. The similarity of nodes in each snapshot was achieved by fusing the latent feature similarity and topological similarity using the L₂ norm. The traditional trend-moving average method was enhanced by assigning a decayed factor to each snapshot, ensuring that the latest snapshot had the greatest weight. The possibility of a connection between nodes at the next step was obtained using the improved trend-moving average method.Results and Discussions Experiments were conducted on four real opportunistic network datasets, ITC, MIT, Infocom05, and Infocom06, where the dataset information included the number of nodes and the collection time of the datasets. AUC (area under the receiver operating characteristic curve) and F1-score were used as evaluation indicators of model performance. The experiments were divided into slicing slot, model complexity, model generalization, comparative experiment, and ablation experiment phases. Based on the experiments on network sample entropy, Graph2vec model training time, TOPSIS comprehensive score, and AUC value under different slicing slots, it was found that when the slicing slots are 8, 10, 3, and 6 minutes for datasets ITC, MIT, Infocom05, and Infocom06 respectively, the comprehensive evaluation score curve reaches a stable trend, and AUC achieves a good effect. Therefore, 8, 10, 3, and 6 minutes were selected as the slicing slots for the ITC, MIT, Infocom05, and Infocom06 datasets. The complexity of the model was compared and analyzed from the perspectives of model complexity, including the proposed model and baseline models, as well as the FLOPs and Params of the models. In addition, the prediction time of the models was compared. The accuracy of the proposed method was tested using ten-fold cross-validation. On the ITC, MIT, Infocom05, and Infocom06 datasets, the average AUC of the proposed method exceeded 0.92, with the average AUC on the Infocom05 dataset exceeding 0.95, demonstrating that the model has good generalization. After removing the node topological similarity module in the model for ablation experiments, the AUC of the proposed model decreased by 1.07% to 3.86%, and the F1-score reduced by 6.12% to 15.48%. This validates the effectiveness of the node topology similarity module. Compared to the baseline methods LSTM (long short-term memory), NESND (network embedding supplementing similarity information in the network domain), E–LSTM–D (encoder–LSTM–decoder), GCN–GAN (graph convolutional network–generative adversarial network), SE–GRU (structure embedded gated recurrent unit neural networks), and GC–LSTM (graph convolution embedded LSTM), the AUC improved by 0.5% to 24.8%, and the F1-score increased by 1.15% to 22.77%. Therefore, the proposed model achieves better performance.Conclusions This study proposes a slicing method based on network sample entropy and model training time, as well as an opportunistic network link prediction model, FS–ET, which is based on feature similarity. The experimental results on four real datasets show that, compared to baseline models, FS–ET achieves better performance.