Temporal correlation-based neural relational inference for binary dynamics

Binary-state dynamics are prevalent in nature, from societal dynamics to dynamical systems in physics. Reconstructing a network structure behind interacting binary-state dynamical systems is essential, as it can facilitate understanding of these dynamical systems and improve the accuracy of predicting dynamical behavior. So far, few works have focused on correlation information in binary-state temporal data to help reconstruct networks. In this study, we propose temporal correlation-based neural relational inference for binary dynamics (TCNRI), inspired by the maximum likelihood estimation of activation events in binary dynamics processes. TCNRI constructs instantaneous correlation features and long-term correlation features by analyzing activation events in the time series data. These features capture the correlation information and help TCNRI reconstruct the network structure. We treat the binary-state dynamical process as a Markov process and use neural networks to reproduce node dynamics based on the reconstructed network structure. We conduct simulations on the classic susceptible–infected–susceptible (SIS) dynamics and Ising dynamics. The results show that TCNRI significantly outperforms baseline models and can accurately reconstruct the network structure for both typical synthetic networks and real networks.