Note: Naji Shajari has transitioned from the institute (alumni).
Lab Rotation: A Toy Model for Visualization of Anticausal Learning in The Case of Two 1D Variables
I did my first lab rotation in the causal inference group at MPI supervised by the group leader, Dr. Dominik Janzing. I have been part of a study by my supervisor and colleagues about the effectiveness of anticausal semi-supervised learning compared to causal scenario. The idea is that P (X = x) will prove helpful in learning P(Y = y|X = x) only when Y → X (read Y causes X), presuming that the ground truth is either Y → X or X → Y . I have managed to implement a computationally efficient toy model using dynamic programming. I also used this opportunity to further develop my skills in Python programming; beforehand I did my usual programming tasks in C++. My contribution to the research has been appreciated by my supervisor up to the level that he decided to include me as a co-author in an article [2].
Master Thesis: A Novel Causal Inference Method For Time Series
I have done my master thesis as part of causal inference group as well. My work was led by Dr. Dominik Janzing and Dr. Michel Besserve. I worked on developing a causal discovery method for time series based on the frameworks which were initiated in our group. And subsequently to apply them to real world problems like building effective brain networks. Based on a preliminary idea by Dr. Besserve, I have extended a causal discovery method for deterministic linear high-dimensional data [3] to address the problem of causal discovery for time series. To be more specific, suppose Xt and Yt are given stationary stochastic time series where there exist a linear time-invariant filter L such that Yt = L(Xt) and the question is whether Xt → Yt or Yt → Xt? By the help of my supervisors I was able to prove some identifiability results and to derive an interpretation of this new method which bridged this work to information-geometric causal inference framework [4]. Despite a successful accomplishment, many questions remain to be answered a few of which are: addressing the existence of confounders, establishing a statistical significance test for this method and tackling the case of multivariate time series, the latter of which we are currently working on.