A Fast and Inherently Nonnegative Latent Factorization of Tensors Model for Dynamic Directed Network Representation

A dynamic directed network (DDN) is becoming pyramidally popular to character complex interactions across various entities. When extracting valuable knowledge from it, a latent factorization of tensors (LFT) model has proven to be effective due to its excellent representation learning ability to a DDN. However, current a LFT model often suffers from constraints in its training scheme and slow convergence. In order to address these issues, this paper proposes a Fast and Inherently Nonnegative Latent factorization of tensors (FINL) model. It adopts two-fold ideas: a) introducing a nonnegative mapping function makes parameters training unconstrained when using a general optimization method, and b) incorporating a momentum method into parameters learning scheme to achieve high convergence rate. Experimental results on four real-world DDNs demonstrate that the proposed FINL model outperforms state-of-the-art models in implementing representation learning to a DDN. This is positively supported by its high prediction accuracy and convergence rate for missing links of a DDN.