The Effect of Incomplete Data on Network Reconstruction

Abstract

An understudied aspect in the field of Network Science is to account for errors in observed data. This leads to networks being studied that may not be fully correct and to potentially unsupported conclusions. In this thesis we focus on Cognitive Social Structures, which use network data that stems from the reports of the members of the network. These reports are notoriously unreliable, and so taking the proper steps to account for errors is even more important. As a way to counteract such errors, this thesis studies Bayesian inference, particularly Variational Inference, to obtain a probability distribution over the network instead of a single ‘true’ network. We do this using the VIMuRe model presented by De Bacco et al.. We present a comprehensive guide on all stages of the algorithm, and build upon the VIMuRe model with the addition of another parameter, income bias. While this addition does not significantly improve the model, we explain the steps required for others to build upon it further.