Mathematical Sciences Seminar - Latent variable models for complex networks: flexible modelling and scalable inference
When:
—
Venue:
Birkbeck Main Building, Malet Street
No booking required
Network data arise in a wide range of applications including social and biological sciences. In many cases different relations and/or a large amount of nodal information is available. Nodal attributes and links are often in strong relation. For example, nodes having similar features are more likely to be connected to each other and vice versa. We introduce a new framework of latent variable network models which combine the information given by heterogenous relational network structures in order to analyse, identify latent traits or groups, visualise network data and predict missing links and nodes. Network data are typically of large size and all the likelihood functions of the models proposed cannot be evaluated analytically. To overcome this problem, we adopt a variational approach to estimation which turns out to improve considerably the computational efficiency without a significant loss of accuracy with respect to other existing methods. The effectiveness of this methodology is demonstrated on the analysis of a wide variety of networks (from small to large networks). The analysis is carried out with the lvm4net package for R (available on CRAN).
Contact name:
Department of Economics, Mathematics and Statistics