FCause: Flow-based Causal Discovery

Tomas Geffner, Emre Kiciman, Angus Lamb, Martin Kukla, Miltiadis Allamanis, Cheng Zhang

Abstract: Current causal discovery methods either fail to scale, model only limited forms of functional relationships, or cannot handle missing values. This limits their reliability and applicability. We propose FCause, a new flow-based causal discovery method that addresses these drawbacks. Our method is scalable to both high dimensional as well as large volume of data, is able to model complex nonlinear relationships between variables, and can perform causal discovery under partially observed data. Furthermore, our formulation generalizes existing continuous optimization based causal discovery methods, providing a unified view of such models. We perform an extensive empirical evaluation, and show that FCause achieves state of the art results in several causal discovery benchmarks under different conditions reflecting real-world application needs.