Machine learning applications in complex systems

Abstract

We observe complex interacting systems on a daily basis. The vast in ternet itself, the network of roads, the network of neurons comprising our nervous system, and so on. Their behaviors have been a matter of inter est for decades. Much of what is understood about the dynamics of these complex systems comes from mathematical estimations and continuous ob servations of their time series. To the same effect, there has been recent developments in machine learning techniques to better understand and pre dict behaviours of these dynamical systems. The main idea behind these techniques is to observe patterns in the data generated from these complex systems, and allow the machine to learn them in order to make certain predictions about the network. In this work, we explore one such novel application of machine learning techniques to unravel some fundamental structure-to-dynamics relations that better help understanding these com plex systems. We classify different types of networks based on their inherent structural differences by training a CNN model on the time-series of a few highest degree nodes. The novelty of our work lies in the observation that using only a limited time-series information of a large network, we make extremely accurate classifications, and prove that with increasing the size of the networks, the number of time-series required remains the same.