Link prediction via nonnegative matrix factorization enhanced by blocks information

Low rank matrices approximations which have been used in networks link prediction are usually global optimal methods and use little local information. However, links are more likely to be found within dense blocks. It is also found that the block structure represents the local feature of matrices because entities in the same block have similar values. So we combines link prediction method by convex nonnegative matrix factorization with block detection to predict potential links using both of global and local information. A probabilistic latent variable model is presented by us and the experiments show that our method gives better prediction accuracy than original method alone (For example, AUC=0.861991 is higher 10% on Karate club network with 5% missing links.).