Title: Multi-source causal discovery from real-world experiments with extended JCI
Authors: Tom Claassen - Radboud University Nijmegen (Netherlands) [presenting]
Abstract: The focus is on an extension of the recently introduced Joint Causal Inference (JCI) framework for causal discovery from multiple data sets. The framework allows us to jointly learn both the causal structure and targets of interventions from statistical independencies in pooled data. Being able to exploit the information and background knowledge from multiple observational and experimental data sets simultaneously provides for a significant improvement in the accuracy and identifiability of the predicted causal relations, while the systematically pooled data also increases the statistical power of independence tests. Previous implementations of JCI were based on powerful SAT solver approaches, which are very robust and flexible, but unfortunately scale poorly, restricting application to relatively small models. This novel adaptation shows how to extend the JCI framework to standard constraint-based algorithms such as FCI+, which makes it possible to handle both soft and hard (perfect) interventions, as well as larger models up to dozens of variables. We will demonstrate the method on a realistic dynamical simulation of interventions on a gene regulatory network, as well as on several real-world data sets.