Short- and long-term temporal network prediction based on network memory

Abstract Temporal networks like physical contact networks are networks whose topology changes over time. Predicting future temporal networks is crucial e.g., to forecast and mitigate the spread of epidemics and misinformation on the network. The classic temporal network prediction problem that aims to predict the temporal network in the short-term future based on the network observed in the past has been addressed mostly via machine learning algorithms, at the expense of high computational costs and limited interpretation of the underlying mechanisms that form the networks. This motivates us to develop network-based models to predict future network based on the network properties of links observed in the past. Firstly, we explore the similarity between the network topologies (snapshots) at any two time steps with a given time lag/interval. We find that the similarity is relatively high when the time lag is small and decreases as the time lag increases. Inspired by such time-decaying memory of temporal networks and recent advances, we propose two models that predict a link’s activity (i.e., connected or not) at the next time step based on past activities of the link itself or also of the neighboring links, respectively. It is observed that in seven real-world physical contact networks, our models outperform in both prediction quality and computational complexity, and predict better in networks that have a stronger memory. We also reveal how different types of neighboring links contribute to the prediction of a given link’s future activity, again depending on the properties of temporal networks. Furthermore, we adopt both models as well as baseline models for long-term temporal network prediction, that is, predicting temporal networks multi-time steps ahead based on the network topology observed in the past. Our models still perform better than baseline models at each step ahead in long-term prediction, and our models have a higher prediction quality in networks with stronger memory. The prediction quality of our SD model decays as the prediction step is further ahead in time and this decay speed is positively correlated with the decay speed of network memory. Finally, networks predicted by various models respectively have a heterogeneous distribution of inter-event time similar to real-world networks, and also the burstiness of inter-event times for individual link