Skip to Main content Skip to Navigation
New interface

Learning causal graphs from continuous or mixed datasets of biological or clinical interest

Abstract : The work in this thesis follows the theory primarily developed by Judea Pearl on causal diagrams; graphical models that allow all causal quantities of interest to be derived formally and intuitively. We address the problem of causal network inference from observational data alone, i.e., without any intervention from the experimenter. In particular, we propose to improve existing methods to make them more suitable for analyzing real-world data, by freeing them as much as possible from constraints on data distributions, and by making them more interpretable. We propose an extension of MIIC, a constraint-based information-theoretic approach to recover the equivalence class of the causal graph from observations. Our contribution is an optimal discretization algorithm based on the minimum description length principle to simultaneously estimate the value of mutual (and multivariate) information and evaluate its significance between samples of variables of any nature: continuous, categorical or mixed. We use these developments to analyze mixed datasets of clinical (medical records of patients with cognitive disorders; or breast cancer and being treated by neoadjuvant chemotherapy) or biological interest (gene regulation networks of hematopoietic stem and precursor cells).
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Friday, June 10, 2022 - 4:03:16 PM
Last modification on : Tuesday, June 28, 2022 - 3:14:13 AM
Long-term archiving on: : Sunday, September 11, 2022 - 7:37:47 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03693627, version 1


Vincent Cabeli. Learning causal graphs from continuous or mixed datasets of biological or clinical interest. Artificial Intelligence [cs.AI]. Sorbonne Université, 2021. English. ⟨NNT : 2021SORUS331⟩. ⟨tel-03693627⟩



Record views


Files downloads