For the past two decades, electronic devices have revolutionized the traceability of social phenomena. Social dynamics now leave numerical footprints, which can be analyzed to get a deeper understanding of collective behaviors. The development of large online social networks (like Facebook, Twitter and more generally mobile communications) and connected physical structures (like transportation networks and geolocalised social platforms) resulted in the emergence of large longitudinal datasets. These new datasets bring the opportunity to develop new methods to analyze temporal dynamics in and of these systems. Nowadays, the plurality of data available requires to adapt and combine a plurality of existing methods in order to enlarge the global vision one has on such complex systems. The purpose of this thesis is to explore the dynamics of social systems using three sets of tools: network science, statistical physics modeling and machine learning. This thesis starts by giving general definitions and some historical context on the methods mentioned above. After that, we show the complex dynamics induced by introducing an infinitesimal quantity of new agents to a Schelling-like model. It results in a chaotic behavior depending on population composition. The third chapter explores the differences between global method and local method for temporal community detection using scientometric networks. The fourth chapter shows the added value of using longitudinal data. We study the behaviors evolution of bike sharing system users and analyze the results of an unsupervised machine learning model aiming to classify users based on their profiles. The last chapter merges complex network analysis and supervised machine learning in order to describe and predict the impact of new businesses on already established ones. We explore the temporal evolution of this impact and show the benefit of combining networks topology measures with machine learning algorithms.