Project Overview

JProGraM is an open-source Java library which can be used for learning the following probabilistic models from data: Bayesian networks, Markov random fields, hybrid random fields, probabilistic decision trees, dependency networks, Gaussian mixture models, and Parzen windows. Along with learning algorithms, some simple inference methods are also supported by JProGraM. JProGraM is free software: it is released under the GNU General Public License. It is not yet as polished as I would like it to be, and there is no proper documentation, but it works pretty well, and it should be especially useful for research purposes. If you would like to contribute to the development of JProGraM, please contact me. Some ideas for extending the project are listed below. The JProGraM project is also hosted at SourceForge.net.

The source code of JProGraM can be downloaded from here (as a gzipped tarball). For convenience, the tarball also includes some other open-source libraries that are called within JProGraM. In particular, along with JProGraM you're going to download (parts of) WEKA, JGraphT, and the L-BFGS implementation provided by the RISO project. The L-BFGS method is called by JProGraM using a wrapper written by Jacob Eisenstein.

For more information concerning the models and algorithms implemented by JProGraM, you can have a look at my Ph.D. dissertation. If you want to refer to JProGraM in a publication, please refer to my dissertation (BibTeX) and/or to this website.

Project Ideas

• Higher-priority projects:


• Graphical user interface. Creating a GUI for JProGraM would be an excellent way of making the library also usable as a stand-alone application for statistical machine learning. Also, in order for a GUI to work smoothly, it is necessary to standardize the format of the datasets used by JProGraM's learning algorithms, since training sets are currently not represented uniformly across the implemented algorithms.


• Command line interface. Creating a CLI for JProGraM would be another way of making the library also usable as a stand-alone application for statistical machine learning. Moreover, it is necessary to standardize the format of the datasets used by JProGraM's learning algorithms, since training sets are currently not represented uniformly across the implemented algorithms.


• Lower-priority projects:


• Inference algorithms. Currently, Markov chain Monte Carlo (in particular Gibbs sampling) is the only inference algorithm supported by JProGraM. Implementing more inference methods (e.g. belief propagation techniques) would be very useful.


• Learning algorithms. Adding more structure/parameter learning algorithms for Bayesian networks and Markov random fields would be great (both in terms of search algorithms and scoring functions).


• Continuous domains. Adding support for learning/inference in continuous domains would be great, e.g. by implementing Gaussian Bayesian networks.