Seminar in applied mathematics and statistics
SPEAKER: Joe Suzuki (Department of Mathematics, University of Osaka).
TITLE: Causal order Identification based on Minimizing Mutual Information among the Noises.
ABSTRACT: This paper considers an improved and extended version of LiNGAM (Linear Non-Gaussian Acyclic Model) that determines the order of variables from dataset so that cause precedes effect when the variables are expressed by a set of linear equations added by noises. The current LiNGAM for more than two variables seeks the causal order in a greedy manner and assumes that no confounder exists. Based on an extended notion of ICA (Independent Component Analysis), this paper proposes to globally search the causal order so that the mutual information among the noises is minimized. Thus, the original LiNGAM is extended so that it can deal with confounders. While in general, in the actual search of $p$ variables, $p!$ mutual information values should be compared, this paper reduces the task to the shortest path problem, and save the computation dramatically. Experiments using artificial and actual data show that the proposed version of LiNGAM (LiNGAM-MMI) shows significantly better performances than the DirectLiNGAM even when no confounder exists, and examine that it executes in realistic times for the actual data sets.
Upcoming events:
Wednesday, March 27 at 11.15: Kathryn Colborn
Friday, March 29 at 13.15: Jevgenijs Ivanovs
Friday, May 3 at 13.15: Mette Asmild
Thursday, May 23 at 14.15: Andrea Macrina