Belief Network (BN)
PowerConstructor
(Other
systems in the package: BN PowerPredictor, Data PreProcessor)
Introduction
This system is based on my three-phase belief network
construction algorthims. It includes a wizard-like user interface and a
belief network construction engine. (A recent technical report about the
algorithms and the system is available for download: report98.ps.gz
or report98.pdf)
- Platforms: 32-bit windows systems on PC (Windows95/98, 2000 and NT).
- Version: 2.2 Beta. (Update History: Last update August 14, 2001)
- Input: A data set and optional domain knowledge (attribute ordering, partial ordering, direct causes and effects -- initial structure, forbidden links, root and leaf nodes). (Continuous values can be detected and discretized.)
- Output: The belief network (including both structure and parameters) of the data set.
Features
User
interface: (Click here to see a canned demo -- a group of screen
shots)
- Wizard-like interface. It gathers necessary input information through 5 simple steps.
- Online help for each step.
- Graphical belief network editor for modifying BN structure after the learning process.
Construction
engine:
- Accessibility. It supports most of the popular desktop database and spreadsheet formats, including: Ms-Access, dBase, Foxpro, Paradox, Excel and text file formats. It also supports remote database servers like ORACLE, SQL-SERVER through ODBC.
- Reusability. The engine is an ActiveX DLL(bnpcapi.dll), so it can be easily integrated into other belief network, data mining or knowledge base systems for windows95/98/NT.
- Efficiency. This engine constructs belief networks by using conditional independence(CI) tests. In general, it requires CI tests to the complexity of O(N4); when the attribute ordering is known, the complexity is O(N2). N is the number of attributes (fields). To get a rough idea about the running time, see the description of some of my experiments. (In practice, the running time is about O(N2) even without node ordering.)
- Reliability. Modified conditional independence test method is used to make the results more reliable when the data set is not large enough.
- Supporting domain knowledge. Complete ordering, partial ordering, direct causes and effects, forbidden links and root & leaf nodes can be used to constrain the search space and therefore speed up the learning process.
- Supporting large data sets. Running time is linear to the number of cases. See the description of experiment 2.
- Reusability of CI test results. A log file which contains CI test results of a particular data set can be created. This information can speed up the construction process greately when the same data set is being analyzed again with different settings.
- Connectivity: The resulting belief network can be exported to other belief network systems. (The current version supports Hugin net format of Hugin Expert A/S, Netica format of Norsys Software Corp and Bayesian Interchange Format version 0.15 -- BIF 0.15.)
- Supporting condensed data sets, which has a 'frequency' fields that contains the number of appearances of the current entry in the data.
What's new in Version 2
- Support more types of domain knowledge, including: forbidden links, root & leaf nodes. Direct causes & effects (initial structure) can now be imported from several popular BN formats.
- Graphical BN editor is added to the system. (Learning procedure: 1. select input data set; 2. input domain knowledge; 3. start BN structure learning; 4. modify the learned structure using BN editor; 5. start parameter (CP tables) learning and export. )
- The system can now learn the parameters of BN as well.
- The system now support condensed data sets.
Download the package now! (including BN
PowerConstructor, BN PowerPredictor and Data PreProcessor)
Contact: Jie Cheng <jcheng@cs.ualberta.ca>
You are visitor number 
since 30 October 1998.