Link prediction is a method to restore the missing information of a network by predicting the actual but unobserved links or possible new links from the observed links. Currently, link prediction is mainly based on deep learning methods of graph neural networks, which compared with rule-based heuristics, can effectively utilize the network topology information, greatly improves the performance of network link prediction and can be applied to a wider range of network types. However, the existing graph neural network-based methods only use the relative position information of nodes in the network, ignore the basic attributes of nodes and the neighboring information of links, and cannot distinguish the importance of different nodes to the formation of links. To addess these limitation, this paper proposes a link prediction method based on graph attention network and feature fusion, the input features of the network are enriched by adding features such as the degree of nodes, the number of common neighbors of links and the maximum degree of common neighbors. This method first extracts the subgraph centered on the target node pair, and then transforms it into a corresponding line graph, where the nodes in the line graph correspond to the links in the original graph, thus fusing the nodes and link information of the original graph into the nodes of the line graph, which improves the effectiveness and interpretability of feature fusion. Meanwhile, the proposed method uses the graph attention mechanism to learn the weights of nodes, which enhances the flexibility of feature fusion. Experimental results on various network datasets show that the proposed method achieves over 90% in terms of AUC and AP, outperforming existing state-of-the-art methods. Moreover, the method maintains more than 80% prediction performance even when there are many missing observed links.
ICP: