Greedy Learning of Causal Structures in Additive Moise Models

Specialeforsvar ved Phillip Bredahl Mogensen

Titel: Greedy Learning of Causal Structures in Additive Noise Models

 Abstract: Drawing causal inference from data is an important, but also very difficult task due to the vastness of the space of directed acyclic graphs. Taking an unpublished article by Jonas Peters and Martin Wainwright as a starting point, we present the Greedy entropy-search, score-based, Greedy search algorithm for causal discovery in Additive Noise Models. Greedily searching the space of graphs reduces the complexity of causal discovery from super-exponential to polynomial. We prove that the Greedy entropy-search recovers the true causal structure of an Additive Noise Model in the population case and illustrate its finite-sample performance by simulation. 

Vejleder: Jonas Martin Peters
Censor:   Pierre Pinson, DTU