On the Temporal Network Analysis With Link Prediction

Literature review shows that much effort has been given to model physical systems involving a large number of interacting constituents. As a network evolves its constituents (or nodes) and associated links would either increase or decrease or both. It is a challenge to extract the specifics that underlie the evolution of a network or indicate the addition and/or removal of links in time. Similarity-based algorithm, Maximum likelihood methods, and Probabilistic models are 3 mainstream methods for link prediction. Methods incorporating topological feature and node attribute are shown to be more effective than most strategies for link prediction. However, to improve prediction accuracy, an effective prediction strategy of practicality is still being sought that captures the characteristics fundamental to a complex system. Many link prediction algorithms have been developed that handle large networks of complexity. These algorithms usually assume that a network is static. They are also computationally inefficient. All these limitations inevitably lead to poor predictions. This paper addresses the link prediction problem by incorporating microscopic dynamics into the matrix factorization method to extract specific information from a time-evolving network with improved link prediction. Numerical experiments in applying static methods to temporal networks show that existing link prediction algorithms all demonstrate unsatisfactory performances in link prediction, thus suggesting that a new prediction algorithm viable for time-evolving networks is required.