Not Logged In

A Model Selection Criteria for Learning Belief Nets: An Empirical Comparison

Full Text: criteria-00ML.ps PS
Other Attachments: criteria-Wrkshp.ps PS
  criteria-LONG.ps PS

We are interested in the problem of learning the dependency structure of a belief net, which involves a trade-off between simplicity and goodness of fit to the training data. We describe the results of an empirical comparison of three standard model selection criteria --- viz., a Minimum Description Length criterion (MDL), Akaike's Information Criterion (AIC) and a Cross-Validation criterion --- applied to this problem. Our results suggest that AIC and cross-validation are both good criteria for avoiding overfitting, but MDL does not work well in this context.

Citation

T. Van Allen, R. Greiner. "A Model Selection Criteria for Learning Belief Nets: An Empirical Comparison". International Conference on Machine Learning (ICML), Stanford University, July 2000.

Keywords: model selection criteria, belief nets, machine learning, probabilistic graphical models, empirical, theoretical
Category: In Conference

BibTeX

@incollection{VanAllen+Greiner:ICML00,
  author = {Tim Van Allen and Russ Greiner},
  title = {A Model Selection Criteria for Learning Belief Nets: An Empirical
    Comparison},
  booktitle = {International Conference on Machine Learning (ICML)},
  year = 2000,
}

Last Updated: August 13, 2007
Submitted by Russ Greiner

University of Alberta Logo AICML Logo