The SING algorithm learns the structure of non-Gaussian graphical models by estimating a conditional independence score matrix. This matrix is based on Hessian information of the joint log-density of a collection of random variables. Given samples from a target density, the algorithm learns a transport map to represent the density and to define a sparse estimator for the score matrix. More information on the algorithm can be found in the preprint.
Ricardo Baptista (MIT), Rebecca Morrison (UC Boulder), Youssef Marzouk (MIT), and Olivier Zahm (INRIA)
E-mails: rsb@mit.edu, rebeccam@colorado.edu, ymarz@mit.edu and olivier.zahm@inria.fr
The SING algorithm is implemented in Python using the TransportMaps package. The algorithm can be found in the SING folder. Scripts to reproduce the figures for all of the examples in the preprint can be found in the remaining folders.