BEANDisco - Bayesian Exact and Approximate Network Discovery

BEANDisco is a software for learning Bayesian network structure from data. More specifically, it can be used to calculate either estimated or exact (for small networks) probabilities of structural features. Currently only arc features are supported.

Features

Currently only K2-scores supported
Partial order types: Supports both bucket orders and parallel bucket orders
Approximated probabilities with MCMC sampling [1]
Exact probabilities with complete enumeration [2]
Supported feature types: arcs
Separate score computation possible

Possible future enhancements

Support "free" nodes outside buckets (useful for space-time optimal exact computation)
Support more scores
Support more features: markov blanket, arbitrary features by DAG sampling
Optimization

Download and Install

BEANDisco is licensed under GNU GPL 3.0. Download the software from here: BEANDisco-1.0.1.tar.gz (updated 30 Nov 2011)

For compilation you should have Boost library installed. To compile the code run in the source code directory:

make

Examples

To see available parameters run:

./beand --help

Compute estimates with maximum in-degree of 3, (maximum) bucket size 4, burn-in period of 1000 steps and 100 samples with 10 steps between samples:

./beand example.dat -m 3 -b 4 -B 1000 -s 100 -S 10

First compute scores in to a file and then use the precomputed scores to estimate arc probabilities:

./beand example.dat -m 3 --score-file example.score
./beand --score-file example.score -b 4 -B 1000 -s 100 -S 10

Compute exact probabilities (generally with exact computation it is recommended to use a bucket size as large as possible with the available memory):

./beand example.dat -m 3 -b 4 --exact

References

[1] T. Niinimäki, P. Parviainen and M. Koivisto. Partial Order MCMC for Structure Discovery in Bayesian Networks. UAI 2011

[2] P. Parviainen and M. Koivisto. Bayesian Structure Discovery in Bayesian Networks with Less Space. AISTATS 2010