In my talk I describe how the ideal workflow for the study of complex systems can be applied to tagging systems and "folksonomies".

Leveraging the unprecedented amount of data about user activity in on-line social communities collected through on-line systems, the researcher can  identify statistical observables aimed to quantify emergent features, i.e. features which can not be simply deduced by the microscopic action of each single user. Such features should be observed in different systems and situations, calling for a deeper, more general understanding or explanation. 

The very existence of such features is nowadays corroborated by observation of regularities in probability distributions and other finer statistical observables.

Formulation of simple, minimal models can suggest basic mechanisms reproducing the observed phenomenology, as well as evidence relation and dependance between different observations (null models are crucial for this). 

But models are not just descriptive. They predict unobserved features, against which models are checked, and, whenever possible, they allow a mathematical, analytical description of the systems in study. 

Finally, the higher level of understanding attained, feeds back in the design and control of on-line systems, closing a virtuous loop of increasing knowledge and exploitation.
In the case of folksonomy systems, measures and models has been focused on global statistics of user tagging activity, as well as on the emerging correlations between tags, revealed by tag co-occurrence network, whose study involves definition and scrutiny of similarity measures, of their effectiveness and meaning, and suggests the existence of a latent semantic structure of concepts.

Exploiting the full tripartite nature of folksonomy systems (tags, users, resources), similar analysis could be performed in order to reveal emergent classifications of resources, or underlying unexpected relations between users.