This is an implementation of the 2008 paper by Airoldi et al describing a model with context dependent pairwise cluster memberships. This project contains three methods of running inference to obtain the parameters pi and B, where pi is the person's block membership vector, and B is the inter-block relationship matrix.

The first method is simple MAP estimation, using Powell's derivative free function maximizer. The second is slow but accurate MCMC, using an adaptive Metropolis-Hasting algorithm using the awesome Python library PyMC. And the third is a fast deterministic approximation algorithm using variational message passing, built from Microsoft Research's great Infer.NET package.

The dataset is generated by MATLAB as is seen in /data. The MAP and MCMC are written in Python. The variational inference is from Infer.NET which is C#.

This software is released under the MIT license.

The MCMC portion of this software relies on PyMC (freely available under the Academic Free License). Please refer to the PyMC license for details.

The variational inference portion of this software relies on Infer.NET (freely available for non-commercial use) which is not included in our software. Even though the license of this software is permissive, Infer.NET's license is not. Please refer to its license for details.

Alex Burnap, Efren Cruz, Xin Rong, Brian Segal