Anagrafe della ricerca

Networks have been used to model infectious disease spreading in different forms. In general, the nodes represent single individuals and the edges describe the social contacts or sexual relationships through which the infection can spread. The detailed structure and temporal evolution of contact networks were considered for long time almost inaccessible, therefore network epidemiology has mostly focused on the analysis of compartment models in the hypothesis of full mixture or on generalized random graphs characterized by a small set of global statistical measures, such as the degree distribution, level of clustering and community structure. However, detailed spatial and temporal data on social contacts are becoming increasingly accessible because of the diffusion of online social networks on mobile phones and through experiments with radio-frequency identification devices (RFID). In this context, an observation (even partial or noisy) of the infection state of individuals can in principle provide rich information that can be used for health-related purposes, such as the determination of infectivity parameters, the identification of the source of the outbreak or a prediction of future spread development. Once a model has been selected to describe the contagion process, the extraction of this information can be posed as a parametric Bayesian problem where the interesting quantities can be thought as either posteriors, maximum likelihood or maximum a posteriori points. Unfortunately, these values are computationally hard to calculate exactly. We propose to leverage techniques from Statistical Physics to compute these quantities approximately. Such development will provide efficient algorithmic tools, derived from the Cavity Method of Statistical Physics and its single instance counterpart, known as Belief Propagation, for epidemic inference and prediction that could prove to be invaluable for strategic use of resources when fighting large epidemics.