Roger Guimera

Roger Guimerà

Department of Chemical Engineering, Universitat Rovira i Virgili, Tarragona, Catalonia and Catalan Institution for Research and Advanced Studies (ICREA), Barcelona, Catalonia
More information about the Lecturer
Roger Guimerà is ICREA Research Professor of Experimental Sciences and Mathematics at Universitat Rovira i Virgili, in Tarragona, Spain. Roger's research is devoted to the development and application of probabilistic and computational tools for the analysis of complex systems and, particularly, of complex networks. During his career, he has: (i) made methodological contributions to the study of complex networks (e.g. identification of communities and roles, Bayesian network inference, and rigorous model comparison), and (ii) used complex network analysis and network models to gain understanding on specific systems (e.g. social communication and collaboration networks, critical infrastructures such as the air transportation system, and biological systems such as metabolism). These contributions have won him the Erdos-Renyi Prize of the Network Science Society in 2012, and the Young Scientist Award for Socio- and Econophysics of the German Physical Society in 2014.
Transitions in probabilistic inference: From recommender systems to equation discovery
 Inferential methods attempt to understand data and make predictions about the word by explicitly formulating generative models and fitting them. By this process, these methods are able to provide insight on mechanisms, and optimally deal with uncertainty and separate structure from noise. Probabilistically, all the information relevant to an inference problem is captured by the posterior distribution, which quantifies the plausibility of models given the data. This posterior distribution has contributions from the likelihood, which reflects the ability of the model to explain the observed data, and the prior, which encodes for previous knowledge about the system. In this talk, we will discuss how the competition between these two terms induces detectability transitions; these are crucial to understand when inference can work and give useful results, and when it is guaranteed to fail. We will discuss these transitions in the context of two machine learning settings: network-based recommender systems and equation discovery.