B1085
Title: Bayesian causal discovery from unknown general interventions
Authors: Alessandro Mascaro - University of Milano-Bicocca (Italy) [presenting]
Federico Castelletti - Università Cattolica del Sacro Cuore (Milan) (Italy)
Abstract: Directed acyclic graphs (DAGs) are often used to represent causal relationships between variables. In this setting, the process of identifying the DAG structure from data is referred to as causal discovery. If only observational data are available, the DAG is identifiable only up to its Markov equivalence class. However, if in addition one uses experimental data, i.e. data in which the generating process has been altered by an external intervention, then it is possible to identify smaller sub-classes of DAGs, known as I-Markov equivalence classes (I-MECs). Different types of interventions modify the causal structures in different ways and, accordingly, imply distinct definitions of I-MECs. Current causal discovery algorithms from experimental data assume that interventions do not modify the parents of the intervened nodes in the DAG, even when the targets of interventions are unknown. The assumption is relaxed by proposing a Bayesian methodology for causal discovery from experimental data arising from unknown general interventions. The contribution includes (i) providing definitions and graphical characterizations of general I-MECs; (ii) developing priors which guarantee score equivalence of DAGs within the same I-MECs and (iii) devising suitable MCMC schemes to sample from the posterior distribution over DAGs and unknown interventions.
