Évènement

Title: Importance weighted variational graph autoencoders for inference in deep latent position block models

Abstract:

In this presentation, I will first show how latent position models can be made compatible with block models for network analysis through the use of deep generative models. I will focus on the binary edge case and introduce a new random graph model along with a variational graph auto encoding strategy for inference. I will discuss the identifiability of the model and explain how model selection can be performed. I will then move to the sparse discrete edge case in the same deep, block compatible, modelling framework. I will show how importance weighted variational inference can strongly improve the inference procedure over the classical variational auto encoding strategy. I will give the importance weights used for sampling and show that, in the limit, the lower bound obtained converges to the integrated log likelihood of the data. Through this presentation, I will emphasise the need to rely on flexible random graph models to obtain relevant loss functions.