Betweenness centrality-based community adaptive network representation for link prediction

摘要

Link prediction is a fundamental problem in biological network analysis, personalized recommendation, network evolution modeling, etc. It aims at discovering links in the network that are unknown, missing, or will be formed in the future. Network representation learning-based link prediction approaches have drawn extensive attention, due to its high efficiency. The previous approaches use random or hyper parameters to select nodes from neighbors or communities when generating walk sequences. However, they do not fully consider the contribution of nodes to the embedded representation and hence impairing the affect the role of community structure in link prediction. To overcome this limitation and utilize community structure, we propose a betweenness centrality-based community adaptive network representation for link prediction method called CALP, which forms network representation by using betweenness centrality to measure the different contribution of community nodes and neighbor nodes for embedding and then applies it to link prediction. CALP first divides the network into communities. Then, it selects a node from the community nodes or neighbor nodes to join the walk sequence by the contribution of the node to embedding. Finally, it generates the corresponding network representation for link prediction. Experiments on realistic networks such as Cora, Citeseer, etc. show that the accuracy of CALP is much better than other approaches.