[Seminar] Structural Equation Modeling to latent causal representation learning for more trustable ML

The next LIVIA seminar will be held on Thursday, July 21st at 12h00 in hybrid mode.

Title: Structural Equation Modeling to latent causal representation learning for more trustable ML
by Prof. Myriam Tami, Paris-Saclay, CentraleSupélec

Abstract:
Structural equation models (SEM) with latent variables (LVs) are used to model relationships between observable and latent variables. We will present an approach for estimating an SEM model with LVs based on its global likelihood maximization by the EM algorithm. We will give the numerical results of this approach on simulated data and show, via an application on real environmental data, how to practically build a model and evaluate its quality. Finally, we apply the approach developed in the context of a clinical oncology trial to study the longitudinal quality of life data. We show that by effectively reducing the data dimension, the EM approach simplifies the longitudinal analysis of quality of life by avoiding multiple tests. Thus, it helps to facilitate the evaluation of the clinical benefit of a treatment.
Then, after introducing some key concepts from the causality field, we will motivate the interest in considering SEM models with LVs in this growing field of research. Indeed, identifying causal relationships between observed variables has drawn much attention in the fields of statistical learning and AI. This area, well known as Causal discovery, now mainly includes a range of approaches that do not consider the presence of LVs and that encounter limitation in handling a huge number of variables. We will see that SEM with LVs can be a response to these limitations and constitute an interesting line of research to explore together.

Bio:
Myriam TAMI (PhD 2016, University of Montpellier, Institut Montpelliérain Alexander Grothendieck, south of France) is an Associate Professor at University Paris-Saclay, CentraleSupélec, MICS lab. Her research works are on AI, Machine Learning, representation learning, causality, and models in the context of complex or heterogeneous data, e.g., multimodal, structured, and unstructured with sometimes latent variables, with uncertainty or weakly labeled. Her publications and research profile can be consulted on her web page or Google Scholar via the following links.
Web page: https://myriamtami.github.io/
Google scholar: https://scholar.google.com/citations?hl=fr&user=kavk5oUAAAAJ