E-swish Regularized Nonnegative Latent Factorization of Tensors for Dynamic Network Link Prediction

A dynamic network is frequently applied to represent a complex interaction system, where a directed weighted link quantifies the direction and strength of a specific interaction between a pair of entities. Generally, it is impossible to observe the full interactions among all entities at each time slot, so there are numerous missing links in a corresponding dynamic network. However, most of existing link prediction models are unable to predict the direction and weight of a missing link simultaneously, which greatly restrict their generality. To address this issue, this paper presents an E-swish Regularized Nonnegative Latent-factorization-of-tensors (ERNL) model, which adopts a third-order incomplete tensor to represent a dynamic network. Its main ideas include: a) designing an E-swish activation function-based nonlinear regularization scheme to improve prediction accuracy; and b) implementing hyper-parameters self-adaptation to achieve high scalability. Empirical studies on four real datasets illustrate that ERNL achieves higher accuracy and computational efficiency than state-of-the-art predictors in predicting the missing directed weighted links of a dynamic network.