Bayesian Network Repository

Contents

• About This Page
• The datasets
• Network formats and Utilities
• Related Sites

Mission

Our in intention is to construct a repository that will allow us empirical research within our community by facilitating (1)better reproducibility of results, and (2) better comparisons among competing approach. Both of these are required to measure progress on problems that are commonly agreed upon, such as inference and learning.

A motivation for this repository is outlined in "Challenge: Where is the impact of Bayesian networks in learning?" by N. Friedman, M. Goldszmidt, D. Heckerman, and S. Russell (IJCAI-97).

This will be achieved by several progressive steps:

Sharing domains. This would allow for reproduction of results, and also allow researchers in the community to run large scale empirical tests.

Sharing task specification. Sharing domains is not enough to compare algorithms. Thus, even if two papers examine inference in particular network, they might be answering different queries or assuming different evidence sets. The intent here is to store specific tasks. For example, in inference this might be a specific series of observations/queries. In learning, this might be a particular collection of training sets that have a particular pattern of missing data.

Sharing task evaluation. Even if two researchers examine the same task, they might use different measures to evaluate their algorithms. By sharing evaluation methods, we hope to allow for an objective comparison. In some cases such evaluation methods can be shared programs, such as a program the evaluates the quality of learned model by computing KL divergence to the "real" distribution. In other cases, such an evaluation method might be an agreed upon evaluation of performance, such as space requirements, number of floating point operations, etc.

Organized competitions. One of the dangers of empirical research is that the methods examined become overly tuned to specific evaluation domains. To avoid that danger, it is necessary to use "fresh" problems. The intention is to organize competitions that would address a specific problems, such as causal discovery, on unseen domains.