BEANDisco - Bayesian Exact and Approximate Network Discovery
BEANDisco is a program for Bayesian learning of Bayesian network structure from data. More specifically, it can generate samples from the posterior and compute the posterior probabilities of structural features.Features
- Supports both ordermodular and modular structure priors.
- Multiple scoring options (i.e. parameter priors): BDeu, K2, LL, MDL, AIC
- Can compute
- posterior probabilities of arcs and
- normalizing constant (marginal likelihood) of the model
- exactly or
- by sampling bucket orders or linear orders
- Sampling methods: MCMC, MC3, AIS
Download
BEANDisco is licensed under GNU GPL 3.0. The current version is:
BEANDisco-2.0.tar.gz (updated 22 Apr 2016)
(Older versions are also available: 1.0.1 (30 Nov 2011), 1.0 (23 Jun 2011))
Compilation requires a compiler that supports C++11 and the Boost library. See README for further compilation and usage instructions.
Algorithms/References
The novel algorithms implemented in BEANDisco are described in the following article:- Structure discovery in Bayesian networks by sampling partial orders . Journal of Machine Learning Research, 2016.
- Annealed importance sampling for structure learning in Bayesian networks . 23rd International Joint Conference on Artificial Intelligence (IJCAI 2013)
- Partial order MCMC for structure discovery in Bayesian networks . 27th Conference on Uncertainty in Artificial Intelligence (UAI 2011)
- Bayesian Structure Discovery in Bayesian Networks with Less Space . 13th International Conference on Artificial Intelligence and Statistics (AISTATS 2010)
Contact
In case of questions, suggestions or general feedback, please contact Teppo Niinimäki.