Synthetic temporal networks are essential to schematize many real systems whose behavior vary in time, from social interactions to biological systems, and for which real data are not always easily collected, being often incomplete or not shareable due to privacy issues. However the generation of realistic temporal graphs is still an open problem. The main issue relies on mimicking both the temporal and the topological properties of the input network, and all their correlations. We propose a novel simple method to explore a temporal network, consisting in decomposing it in its building blocks, namely local temporal neighborhoods of each node with short memory. We then use them to generate a new network from scratch. Basically, the essential information that we use from the original graph to build the new one concerns the behavior of each node in the short time distance, i.e. which connections it creates, eliminates, or maintains, given the connections in the few previous time steps. We thus generate a new pattern of behavior by preserving the short-term temporal correlation of each node.
Not only our method can generate real interaction patterns, but it is also able to capture the intrinsic temporal periodicity of the network and to generate temporal graphs with an execution time lower of multiple orders of magnitude with respect to other similar models.