Talk announcement of Prof Jan Lemeire February 16th room…
title: “Causal structure learning under unfaithfulness.”
abstract:” Inferring from experimental data the causal relations among variables is a scientific important but challenging task. One of the approaches relies on the probabilistic independencies between variables. Algorithms have been designed that allow the construction of the causal model of the system. Although powerful, one of the fundamental assumptions – namely faithfulness – will be rarely met in practica. Faithfulness states that all independencies found in the data are a result of the causal structure and not from specific parameter configurations. The importance of this property for causal learning is clear: if indeed all independencies come from the causal structure, the independencies provide direct evidence on the causal relations. Unfaithfulness will however appear, as a result of specific parameterizations (e.g. deterministic relations) or as a result of the limited sample size which makes that some dependencies are not measurable (called near-to-unfaithfulness). We will discuss the different cases of unfaithfulness that disrupt the causal learning, what we can do about it and how to deal with limited data. Furthermore it seems to investigate the types of unfaithfulness cases that appear in the MassCausal data. Understanding it’s nature might help the causal learning process.”
