Optimal time window in time series

In this semester research project within the chairs of Systems Design (ETH), we followed a two-fold approach. First, time-stamped interactions are aggregated according to a specified time window size, leading to a sequence of static aggregated networks. Second, relations are extracted from each aggregated network. The main challenge is that relations between system elements may change over time and the choice of the time window size is crucial to distinguish meaningful topological changes from random fluctuations. We overcome this problem by identifying regularities in the sequence of aggregated networks at different time windows, specifically by measuring the similarity between consecutive networks. This allows us not only to find the optimal time window size but also to identify precisely the timescales at which relations between agents change.