Archives

INFO : Devenir expert en branches avec Git

10 février 2026 @ 13 h 15 min – 16 h 15 min –

Campus Triolet, Pl. Eugène Bataillon, Bat 36, TD 36.406
Céline Mandier – Institut de Science des Données de Montpellier

Venez assister à un atelier pratique pour maîtriser la gestion des branches avec Git, un outil essentiel pour une collaboration efficace en recherche. Vous apprendrez à créer, gérer et utiliser les branches selon les bonnes pratiques pour optimiser le travail en équipe.

INFO, Git, Branches

Causal Inference from Longitudinal Data: Latent Adjustment Variables and Counterfactual Prediction

10 février 2026 @ 14 h 00 min –

Room 02.124, Building 5, St Priest campus

Machine Learning in Montpellier, Theory & Practice – Myriam Tami (Univ. Paris-Saclay / CentraleSupélec)

The presented works focus on causal inference from longitudinal data, with two complementary axis. The first addresses settings in which some adjustment variables are latent or unobserved, while the second targets the estimation of long-term treatment effects. The first contribution investigates how to account for unobserved adjustment variables using a probabilistic generative model. We propose a causal dynamic variational autoencoder, which learns a latent representation intended to substitute for the missing variables. This approach relies on a conditional Markov assumption linking the latent space to the unobserved variables. The model is trained by maximizing a regularized conditional likelihood, in order to handle imbalances in the representation space and to enhance the stability of temporal inference. The second contribution aims at improving long-term counterfactual prediction. The goal is to learn a compact latent representation capable of summarizing the relevant information from the historical data to predict outcomes under a given hypothetical treatment. A contrastive approach is used to capture long-term dependencies, and is further regularized to ensure invertibility of the latent representation and preservation of confounding and static information, using mutual information criteria. Selection bias is also addressed by minimizing the dependence between the current representation and the future treatment assignment. The proposed methods are evaluated on both synthetic and real-world datasets.

Machine Learning in Montpellier, Theory & Practice