Aller au contenu. | Aller à la navigation

Université
Facultés
Études et Formations
Recherche
Service à la société
International

Abstracts Session 4

16H15-17H30 SESSION 4: STATISTICS METHODOLOGY

1) Lise Léonard (UCLouvain) and Eugen Pircalabelu (UCLouvain): Aggregating estimators from distributed sources. A journey from estimation to model selection (Part 1)
2) Ensiyeh Nezakati Rezazadeh (UCLouvain): Aggregating estimators from distributed sources. A journey from estimation to model selection (Part 2)

Abstract:

The need to analyse in parallel multiple datasets or to perform local, separate analyses is nowadays more and more stringent due to security and privacy concerns. More than that, for problems that depend on unknown tuning parameters it is standard to fit models in parallel on grids of different values, in the hope of identifying one such value that empirically performs best.

This talk will focus on methods to aggregate estimators of unknown coefficients governing different data generating processes, where the estimation is done in parallel. Rather than selecting one estimator, the main feature of our procedures is that it pools together information from multiple estimators to obtain a final, aggregated one. The first part of the talk will start from an ℓ1 penalized, high-dimensional linear model and will propose strategies to aggregate regression coefficients, while the second part of the talk will focus on the case of ℓ1 penalized, high-dimensional graphical models.

Throughout the presentation, real and simulated numerical examples illustrate the usefulness of the methods in practice and theoretical guarantees are investigated.