|
Abstract:
|
Variational inference is gaining popularity in the analysis of network data. However, several challenges in the implementation of the algorithm exist. One of the key challenges concerns the choice of variational family. Furthermore, the distribution of the observed data tends to be misspecified. In this presentation we describe a new algorithm, VDMIX, to provide robust and efficient estimates of the parameters. We also present results on the computational complexity of the algorithm and provide comparisons with the traditional variational algorithm. We apply our results to obtain estimates of privacy risk in networked data sets.
|
