Zamanla değişen dinamik ağlar için uygun zaman penceresi boyu seçimi

Proper dynamic network extraction is a prominent problem for timely evolving systems' modelling. Dynamic network modelling of timely evolving complex systems allow to discover emerging properties of real-world facts. The main issue of such modelling is determining the proper time intervals for each member of network. The proper window size should be detected for extracting the most informative and least noisy time series of snapshot features. Existing solutions suffer from using only network topological properties as snapshot features, applying subjective methodologies needed user-dependent data labeling, and extracting snapshots with equal sized windows. We propose not only a new network similarity based compression ratio measuring the properness of studied window size, but also a window aggregation strategy allowing to extract a more informative dynamic network with less noisier structure by using the similarity metric and its statistical significance. The results on Enron, Haggle Infocom and Reality Mining data sets reveal that the proposed compression ratio is more effective for finding best window size than baseline, and aggregation strategy allow to capture important dynamic events which might be hidden in noise when using constant windows. Moreover, the statistical results show that the general topological characteristics of the network is mostly not affected after aggregation process. Finally, as a complementary result, aggregation process can be used in determining proper time interval for modelling.