Seminar@SystemX/Institut DATAIA - Charles Bouveyron

Published on May 2, 2023 Updated on May 2, 2023

on the May 10, 2023

02:00pm - 03:00pm
CentraleSupélec (Bât. Bouygues) - Amphithéâtre Peugeot (sc.046) and online
Charles Bouveyron (Director of the 3IA Côte d'Azur Institute) will lead the seminar co-organised by SystemX and the DATAIA Institute on "Deep latent variable models for unsupervised learning from interaction data" at CentraleSupélec.

This event is organised in collaboration with the IRT SystemX.

Topic: "Deep latent variable models for unsupervised learning from interaction data

In this talk we will focus on the problem of statistical learning with interaction data. This work is motivated by two real-world applications: modelling and clustering of social networks, on the one hand, and pharmacovigilance data, on the other. To this end, we have developed two model-based approaches. First, we propose the Deep Latent Position Model (DeepLPM), an end-to-end generative clustering approach that combines the widely used latent position model (LPM) for network analysis with a convolutional graph network (CGN) encoding strategy. A novel estimation algorithm is introduced to integrate explicit optimisation of posterior clustering probabilities via variational inference and implicit optimisation using stochastic gradient descent for graph reconstruction. Second, for the pharmacovigilance problem, we introduce a latent block model for dynamic co-clustering of count data streams with high sparsity. We assume that the observations follow a time- and block-dependent mixture of Poisson distributions with zero swelling, which combines two independent processes: a dynamic mixture of Poisson distributions and a time-dependent scattering process. To model and detect abrupt changes in the dynamics of cluster membership and data sparseness, the mixing and sparseness proportions are modelled by systems of ordinary differential equations. Model inference is based on a novel variational procedure whose maximisation step trains recurrent neural networks to solve the dynamical systems. Numerical experiments on simulated data sets demonstrate the effectiveness of the proposed methodologies for both problems.

More information