Detecting Low-Complexity Confounders from Data

Maria Virginia Ruiz Cuevas (1) , Nataliya Sokolovska (2) , Pierre-Henri Wuillemin (3) , Jean-Daniel Zucker (4)

1 IRIC - Institut de Recherche en Immunologie et en Cancérologie [UdeM-Montréal]
2 SU - Sorbonne Université
3 DECISION
4 IRD - Institut de Recherche pour le Développement

Statistical dependencies between two variables X and Y indicate that either X causes Y , or Y causes X, or there exists a latent variable Z which influences X and Y. In biology and medicine, an important problem is to find genetic or environmental unobserved causes of phenotypic difference between individuals. In this contribution, we introduce a novel approach to identify unobserved confounders in data. The proposed method is based on the state-of-the-art 3off2 causal network reconstruction algorithm, and on an evidence for a direct causal relation represented by purity of con-ditionals. The proposed method is implemented in Python, and it will be publicly available shortly. We discuss the results obtained on a real biomedi-cal dataset.

Domaines: Intelligence artificielle [cs.AI]; Apprentissage [cs.LG]; Machine Learning [stat.ML]

Maria Virginia Ruiz Cuevas, Nataliya Sokolovska, Pierre-Henri Wuillemin, Jean-Daniel Zucker. Detecting Low-Complexity Confounders from Data. ICML / IJCAI / AAMAS FAIM'18 Workshop on CausalML, Jul 2018, Stockholm, Sweden. ⟨hal-01858403⟩