DYNAMIC NETWORK LINK PREDICTION BASED ON LATENT FACTORIZATION OF TENSORS

Dynamic network link prediction is a fundamental task of dynamic network analysis. Existing dynamic network link prediction models often not consider the direction and weight of missing dynamic links. Therefore, this paper proposes a regularized non-negative latent factorization of tensors (RNL) model. This model built a non-negative learning objective incorporating elastic net regularization and linear biases based on latent factorization of tensors, and designed an optimization parameters learning scheme based on single latent factor-dependent and non-negative, multiplicative update (SLF-NMU) rule. Empirical studies on two dynamic network datasets demonstrate that the proposed RNL model is able to predict the missing directed and weighted links of a dynamic network accurately and efficiently.