A repository holding the code implementing the methods of
"SparsityBoost: A New Scoring Function for Learning Bayesian Network Structure" by Eliot Brenner and David Sontag
Presented at Uncertainty In Artificial Intelligence 2013, Friday July 12, Bellevue, Washington, USA
Published in Uncertainty in Artificial Intelligence: Proceedings of the Twenty-Eighth Conference (2013).
To install the c/ part of the library (the only part needed for learning networks) first type:
from the c/ directory. This should create the bscore executable file in your c/ directory.
The library exists in two independent parts, in different languages and with different purposes: the parts under c/, in C/C++, are for learning networks, and the parts under python/monteCarloBeta are for computing the values of the CDF of the Mutual Information ("beta") needed for the interpolation.
We recommend starting by trying to learn the structure of a network, by assuming that eta (a lower bound for the edge strength) takes one of the following values:
0.04, 0.02, 0.01, 0.005, 0.001
In this case, the user only needs to run the C/C++ parts of the library. A typical invocation of the main method, ../c/bscore, looks like this:
../c/bscore -bic -nodes 37 -maxpa 4 -data ../data/synthetic_examples/experiments/0/alarm1000.dat -mod_out ../results/experiments/0/sparsity_eps005/model1000.mod -edge_scores .005
The .dat file consists of raw data, in the format of a space-separated file. Each row holds one observation which is a line of 0s and 1s, one column for each variable in the system. The .mod file is in GOBNILP Score File Format, described on p. 11 of the GOBNILP manual:
In particular, the first line is the number of variables in the network, and what follows are the local parts of our score corresponding to a pair of edges (on a line by themsevles) and possible parent sets (at the beginning of the following lines with the scores).
In order to use this mod file to actually learn the nework, you can use GOBNILP or any other system that learns network structure from local scores. For example, using gobnilp1.3:
gobnilp1.3/bin/gobnilp ../results/experiments/0/sparsity_eps005/model4400.mod > ../results/experiments/0/sparsity_eps005/output4400.txt
The script perl/run_experiments.pl reproduces the experiments in the UAI paper.
Reproducing UAI 2013 Results:
In order to run perl/run_experiments.pl you have to have
under the perl directory.
In run_experiments.pl, comment in/out the sed commands which will work on your OS. These lines are labelled "OSX" or "Linux". In the perl directory itself, you have to have the GOBNILP settings file gobnilp.set, which comes with this repository. The sed commands will copy this GOBNILP settings file to the appropriate locations, such as../results/experiments/i/sparsity_eps005/, etc.
The data for the experiment can be downloaded from the following location:
The archive file contains a structure like this:
so that if you untar it directly in the toplevel ("SparsityBoost") folder of the project, the paths will be what the run_experiments.pl script expects.
Computing the CDF of the MI ("beta")
The reason for computing your own InterpolationData files is if you wish to learn a network assuming a minimal edge strength ("epsilon") other than the four values cited above. Use the class betaTable.py in MonteCarlo/montecarlo/. Detailed instructions to follow.
Required for the second step of learning the network and for reproducing our results from the UAI paper:
Gobnilp1.3 and its dependency Scip
General installation and usage instructions for gobnilp1.3 are given in the GOBNILP manual linked above.
run_experiments.pl has been tested on OSX 10.8 and Ubuntu Linux.
SparsityBoost is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation
SparsityBoost is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with SparsityBoost; see the file gpl-3.0.txt. If not, write to the Free Software Foundation, 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.