Monitoring user-system interactions through graph-based intrinsic dynamics analysis

Monitoring the evolution of user-system interactions is of high importance for complex systems and for information systems in particular, especially to raise alerts automatically when abnormal behaviors occur. However current methods fail at capturing the intrinsic dynamics of the system, and focus on evolution due to exogenous factors like day-night patterns. In order to capture the intrinsic dynamics of user-system interactions, we propose an innovative graph-based approach relying on a novel concept of time. We apply our method on two large real-world systems (the Github.com social network and the eDonkey peer-to-peer system) to automatically detect statistically significant events in a real-time fashion. We finally validate our results with the successful interpretation of the detected events.