Cold-start link prediction integrating community information via multi-nonnegative matrix factorization

Cold-start link prediction has attracted much attention as a sub-problem of link prediction recently. However, due to the influence of some isolated nodes existing in the network, the network structure is disconnected, so that the existing methods cannot realize the task of link prediction well. Therefore, how to excavate and fuse some available information from the network data to help complete the link prediction is the key to solve this problem. In this paper, we propose a multi-nonnegative matrix factorization model that implements the prediction of missing edges of isolated nodes in the overall disconnected state of the network structure. Through several methods, three global and local attribute information, namely the community membership information of the node attributes, the attribute similarity between the nodes, and the partial first-order structure characteristics existing among the nodes, are extracted on network. Then, using the proposed new model, the cold-start link prediction problem on the structured disconnected network is finally solved by integrating the three kinds of information from multiple perspective. Extensive experiments demonstrate that our proposed method performs better than state-of-the-art methods when solving the cold-start link prediction problem.