Code for Variational Inference for Sparse and Undirected Models by John Ingraham and Debora Marks, ICML 2017.

This software supports inference of discrete undirected graphical models (UGMs) with pairwise interactions. It provides implementations of standard L1-regularized approximate Maximum Likelihood methods as well as Persistent VI, an algorithm that learns a variational approximation of the posterior of a discrete UGM. It also implements non-centered reparameterizations of sparsity-inducing scale-mixture priors, such as the Horseshoe, to do soft variable selection among the interactions ("Fadeout" from the paper).

The ising subdirectory contains MATLAB code with MEX subroutines for inferring Ising models.

Installation. To setup the Mark Schmidt's mex routines for approximate MAP inference baselines, run ising/external/L1General/mexAll.m and ising/external/minFunc/mexAll.m in MATLAB.

The MATLAB script ising/generate_ising_analysis.m can be run to create similar experiments as those in the paper. The results may differ some due to randomly sampled synthetic parameters and data.

Credits This code uses Mark Schmidt's excellent minFunc and L1General packages for approximate MAP inference which are included in ising/external.

The potts subdirectory contains standalone C code for inferring Potts models as well as MATLAB code for analyzing the results and generating synthetic datasets.

Installation To compile the C code, see instructions in potts/pvi/.

Examples The MATLAB scripts potts/generate_potts_experiments.m and potts/generate_protein_experiments.m can be run to recreate the potts experiments from the paper.

Credits The C code for approximate MAP inference of Potts models uses Naoaki Okazaki's C implementation of L-BFGS as well as a C implementation of the Mersenne Twister