B1177
Title: Bayesian learning of directed networks from interventional experimental data
Authors: Federico Castelletti - Università Cattolica del Sacro Cuore (Milan) (Italy) [presenting]
Stefano Peluso - Università degli Studi di Milano Bicocca (Italy)
Abstract: Directed Acyclic Graphs (DAGs) provide an effective framework for learning causal relationships between variables in multivariate settings. Under pure observational data, DAGs encoding the same conditional independencies cannot be distinguished and are collected into Markov equivalence classes. In many contexts, however, interventional data supplement observational measurements that improve DAG identifiability and enhance causal effect estimation. A general Bayesian framework is proposed for multivariate data partially generated after stochastic interventions. The method provides an effective prior elicitation procedure for DAG-model parameters and leads to a closed-form expression for the DAG marginal likelihood. The model is specialized to Gaussian DAGs, and asymptotic properties of DAG estimation are established in terms of posterior ratio consistency. The theoretical results are validated in simulation and are implemented on synthetic and biological protein expression data in a Markov chain Monte Carlo sampler for posterior inference on the space of DAGs.
